fix(gmm): use log-sum-exp in estimate_log_prob_resp to prevent overflow

[schliffen]

Fixes #442

Problem

estimate_log_prob_resp computed the per-sample log-normaliser as:

ln(Σ exp(xᵢ))

For tight Gaussians (small covariance, observation near the mean) the weighted
log-probabilities xᵢ can exceed ~709, causing f64::exp() to overflow to inf.
The subsequent ln(inf) produces an infinite log_prob_norm, and every log_resp
becomes -inf, collapsing all responsibilities to NaN / zero and destabilising EM.

Fix

Replace with the log-sum-exp trick:

max + ln(Σ exp(xᵢ − max))

where max = max_i(xᵢ) is computed independently per sample row via
map_axis(Axis(1), ...). After the shift the largest exponent is 0, so
exp() never exceeds 1.

Proof of equivalence:

max + ln(Σ exp(xᵢ − max))
= max + ln(Σ exp(xᵢ) · exp(−max))     [exp(a−b) = exp(a)·exp(−b)]
= max + ln(exp(−max) · Σ exp(xᵢ))     [factor out constant exp(−max)]
= max + ln(exp(−max)) + ln(Σ exp(xᵢ)) [ln(a·b) = ln(a)+ln(b)]
= max + (−max) + ln(Σ exp(xᵢ))        [ln(exp(−max)) = −max]
= ln(Σ exp(xᵢ))                        ∎

Both the old and new code operate per-row (Axis(1)); only the intermediate
numerical range changes.

Regression test

test_large_weighted_log_prob_numerical_stability builds a 70-feature,
1-component GMM with covariance 1e-10 · I. A point at the cluster mean
produces a log-Gaussian probability of ~741 — above the f64 overflow
threshold of ~709.78. The test asserts log_prob_norm is finite and
responsibilities sum to 1.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 77.56%. Comparing base (60382a3) to head (5407c0b).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #443   +/-   ##
=======================================
  Coverage   77.56%   77.56%           
=======================================
  Files         106      106           
  Lines        7585     7585           
=======================================
  Hits         5883     5883           
  Misses       1702     1702           

☔ 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.