feat: add kiln-benchmarks crate

[joeycarpinelli]

Summary

Implements the full benchmark suite described in docs/design/git-kiln-benchmarks.md.

What's included

New crate: crates/kiln-benchmarks

A publish = false workspace member containing 7 Criterion benchmarks that measure Git's object store performance under Kiln's build cache access patterns. All repos are bare, created in tempfile::TempDir, with no remotes configured.

#	Benchmark	What it measures
1	bench1_cache_lookup	Hit/miss ref lookup at N = 100 / 1k / 10k
2	bench2_ingestion	Write 4 MB rlib + metadata; compression level 0 vs 6; repo size
3	bench3_materialization	Restore cached outputs to FS; odb vs fs timing split
4	bench4_deduplication	Object count after two builds, one crate changed
5	bench5_ref_namespace	Enumerate refs at 1k / 10k / 100k; lookup hit/miss
6	bench6_fetch_simulation	Transfer size estimates + local ingestion timing
7	bench7_gc	Repo size before/after git gc, pack files, binary delta exclusion

crates/kiln-benchmarks/README.md — serves as the results document. Contains benchmark descriptions, scale targets, pre-populated results tables (to be filled in after a run), and a findings/recommendations section.

Config changes

• .config/release-please-config.json — adds kiln-benchmarks with publish: false
• .config/release-please-manifest.json — adds initial version 0.0.0
• .github/workflows/CD.yml — publishes only git-kiln (not the bench crate); adds a benchmarks job that runs cargo bench and uploads target/criterion/ as a 90-day artifact

Verification

cargo build --package kiln-benchmarks --benches  # clean, no warnings
cargo test --workspace                            # all tests pass
# Quick smoke test (cache lookup at N=100):
cargo bench --package kiln-benchmarks --bench kiln -- bench1_cache_lookup/hit/100

The smoke test produced ~21 µs mean latency on an M-series Mac — well within the 5 ms target.