feat: add full-loop fixed-factor Brunet APIs

[LCGaoZzz]

Summary

Adds first-class full-loop fixed-factor Brunet/KL APIs for Spatial EcoTyper training and projection.

Root cause

The downstream LOOCV workflow previously crossed the Python/PyO3 boundary once per NMF iteration and rebuilt invariant transposes and sums repeatedly. A formal 20-repeat, 6-sample workflow performs many projection solves, so that call overhead and repeated setup dominated runtime and made thread budgeting difficult to attest.

Changes

• add fixed_h_brunet and fixed_w_brunet public Python APIs;
• execute the complete loop in one GIL-released Rust call;
• cache fixed-factor transposes/sums and expose a scoped per-call Rayon thread budget;
• preserve R NMF Brunet semantics: float64, epsilon floor every 10 iterations, connectivity checks, and strict stable > stopconv stopping (default first stop at iteration 420);
• attest fixed-factor immutability with SHA-256 and return convergence/KL/runtime metrics;
• fail closed for invalid factors and infinite KL, including V > 0 && WH == 0;
• bump package version to 0.2.0 and add tests plus a reproducible benchmark.

Validation

• cargo test --release --manifest-path rust/Cargo.toml --lib: 7 passed
• python -m pytest tests -q: 26 passed
• 420-iteration random-matrix comparison against the legacy single-step loop: elementwise identical with 1 and 4 threads
• downstream live parity against R NMF 0.30.1: fixed-H/fixed-W at 1, 10, and connectivity-stopped iterations passed
• benchmark, V=300x2000, rank=8, 420 iterations, 8 threads: fixed-W 4.01x; fixed-H 0.94x (projection is the repeated hot path)

Downstream integration: PrimorDecode/omicos-admin#454.