[Perf #478] GaussianNB Digits predict: vectorize class likelihood evaluation

Parent: #478

Canonical v2 Digits is a very clean phase-localized loss: Flow fit is faster (1.122 ms vs sklearn 1.405 ms) but Flow predict is much slower (1.304 ms vs sklearn 0.538 ms), producing ~0.801x end-to-end despite exact score parity.

Profile per-class/per-feature Gaussian log-likelihood evaluation, repeated log/div operations, variance/mean loads, output reductions, memory layout and temporary allocation. Audit whether class constants can be precomputed during fit and whether prediction can be fused/vectorized.

Acceptance: prediction-stage profile; precompute-vs-runtime operation counts; batch-size and feature/class scaling curves; retained exact canonical score parity and learned-state parity for priors/means/variances; predict >=1.05x sklearn on Digits while preserving Iris's large win.

Start by locating the GaussianNB Digits prediction path and run the canonical v2 benchmark while profiling per-class/per-feature Gaussian log-likelihood evaluation. Compare precomputed versus runtime operation counts and measure batch-size and feature/class scaling. Done means exact score and learned-state parity, prediction at least 1.05x sklearn on Digits, and Iris's large win is preserved.