Benchmark improvements

[yebai]

Fresh-buffer benchmark setup

• In DynamicPPL.TestUtils.AD.run_ad, use setup = deepcopy($params) for both the primal and gradient @be calls so each Chairmarks sample starts from a fresh input buffer instead of reusing the same vector across samples. Matches Mooncake's bench harness; setup runs outside the timed window, so the copy isn't measured.

Plus some minor tweaks.

This appears to have a noticeable impact on log-density evaluation for tiny models, first noticed in #1363: e.g., simple assume observe（linked = true) logdensity improves from ~20 ns to ~4 ns. In practise, these tiny nanosecond overheads won't matter at all, so the improvement here is only for perfectionists.

EDIT: likely LogDensityAt -> logdensity_internal reduced logdensity's overhead by 20 ns in #1363

[github-actions]

DynamicPPL.jl documentation for PR #1396 is available at:
https://TuringLang.github.io/DynamicPPL.jl/previews/PR1396/

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.30%. Comparing base (90a74c3) to head (97456ec).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1396   +/-   ##
=======================================
  Coverage   82.30%   82.30%           
=======================================
  Files          50       50           
  Lines        3543     3543           
=======================================
  Hits         2916     2916           
  Misses        627      627           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Benchmarks @ 97456ec

Performance

Performance Ratio:
Ratio of time to compute gradient and time to compute log-density.
Warning: results are very approximate! See benchmark notes for more context.

===================================================================================================
                                               eval                       gradient                 
                                            ----------  -------------------------------------------
Model                        dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
---------------------------------------------------------------------------------------------------
Simple assume observe*         1     false     5.87 ns       10.47    1200.34       29.01      6.31
Simple assume observe*         1      true     24.2 ns        2.50     323.68        7.03      0.62
Smorgasbord                  201     false     6.39 μs       69.19     126.55        6.36      8.90
Smorgasbord                  201      true     8.82 μs       62.60     124.93        5.39      5.93
Loop univariate 1k          1000     false     19.3 μs      972.11     282.24        7.36      6.14
Loop univariate 1k          1000      true     21.0 μs     1409.72     264.43        6.94      5.79
Multivariate 1k             1000     false     22.5 μs      377.58      80.88        8.98      2.59
Multivariate 1k             1000      true     19.9 μs      280.74      67.13       10.89      2.46
Loop univariate 10k        10000     false    189.0 μs    11080.76     297.10        7.28      5.83
Loop univariate 10k        10000      true    206.0 μs    11127.88     281.18        7.23      5.64
Multivariate 10k           10000     false    202.0 μs     5032.72      89.22       10.99      2.13
Multivariate 10k           10000      true    203.0 μs     5306.03      90.37       11.01      2.16
Dynamic                       15     false     1.42 μs         err      41.80       13.89     10.79
Dynamic                       10      true     1.93 μs        1.91      57.01       17.27     18.37
Submodel*                      1     false      7.1 ns        8.22    1026.07       24.14      4.98
Submodel*                      1      true     7.42 ns        8.20    1081.32       22.97      5.00
LDA                           12      true     22.2 μs        0.45       1.97       34.83       err
===================================================================================================

Rows marked * have t(logdensity) below about 100 ns; their ratios can be dominated by timer floor, fixed overhead, and run-to-run variation. For those rows, raw t(grad) is more meaningful than t(grad)/t(logdensity).

Main @ 90a74c3

==================================================================================================
                                              eval                       gradient                 
                                           ----------  -------------------------------------------
Model                       dim    linked      primal     FwdDiff    RvsDiff    Mooncake    Enzyme
--------------------------------------------------------------------------------------------------
Simple assume observe         1     false     5.19 ns       10.56    1146.17       31.66     10.82
Simple assume observe         1      true     18.6 ns        2.87     443.41       16.46      3.40
Smorgasbord                 201     false     8.94 μs       57.80      90.96        8.51      6.85
Smorgasbord                 201      true     19.9 μs       30.03      52.26        4.27      2.88
Loop univariate 1k         1000     false     51.1 μs      383.49     106.69        3.66      2.77
Loop univariate 1k         1000      true     50.6 μs      517.45     105.61        3.83      2.84
Multivariate 1k            1000     false     44.5 μs      238.08      37.92        4.98      1.93
Multivariate 1k            1000      true     44.3 μs      231.94      38.61        5.64      1.89
Loop univariate 10k       10000     false    227.0 μs    11755.35     244.13        6.30      5.54
Loop univariate 10k       10000      true    240.0 μs    13140.37     228.16        6.29      5.29
Multivariate 10k          10000     false    257.0 μs     6982.18      67.97       10.59      1.85
Multivariate 10k          10000      true    253.0 μs     6730.33      69.64       10.39      2.17
Dynamic                      15     false     2.43 μs         err      37.39       12.31     10.29
Dynamic                      10      true     3.64 μs        1.79      50.33       11.25     17.49
Submodel                      1     false     5.24 ns       20.50    1812.57       60.90     10.06
Submodel                      1      true     5.24 ns       22.08    2269.81       54.39     11.67
LDA                          12      true     40.8 μs        0.55       1.88       21.94       err
==================================================================================================

Environment

Julia Version 1.11.9
Commit 53a02c0720c (2026-02-06 00:27 UTC)
Build Info:
  Official https://julialang.org/ release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 4 × AMD EPYC 7763 64-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-16.0.6 (ORCJIT, znver3)
Threads: 1 default, 0 interactive, 1 GC (on 4 virtual cores)

[sunxd3]

looks good as far as I can tell with one thing to double check