feat: extend accuracy report to cover wavekat-zh backend

[wavekat-eason]

Summary

• Reference probabilities in tests/fixtures/reference.json are now keyed by (backend, file), so multiple Smart Turn variants can share one fixture file. Older flat entries still load — backend defaults to pipecat via serde.
• scripts/gen_reference.py scores both upstream Pipecat v3.2 and the WaveKat zh fine-tune (wavekat/smart-turn-ONNX, zh subfolder) and writes one entry per (backend, clip) pair.
• crates/wavekat-turn/tests/accuracy.rs gains a wavekat module gated on wavekat-smart-turn that loads the zh variant once and contributes rows for the zh_*.wav fixtures. Shared load_wav_f32 / raw_prob helpers lifted out of the pipecat module.
• make accuracy now builds with --features wavekat-smart-turn; the zh weights are pulled from HuggingFace on first run and cached under $HF_HOME/hub/.

Result

make accuracy (v0.0.8):

Backend	Clip	Python	Rust	Diff	Status
pipecat	silence_2s.wav	0.9870	0.9870	0.0000	PASS
pipecat	speech_finished.wav	0.9849	0.9849	0.0000	PASS
pipecat	speech_mid.wav	0.0477	0.0426	0.0051	PASS
pipecat	zh_speech_finished.wav	0.9865	0.9865	0.0000	PASS
pipecat	zh_speech_finished_short.wav	0.9823	0.9843	0.0019	PASS
pipecat	zh_speech_mid.wav	0.0717	0.0717	0.0000	PASS
wavekat-zh	zh_speech_finished.wav	0.7660	0.7498	0.0162	PASS
wavekat-zh	zh_speech_finished_short.wav	0.8751	0.8751	0.0000	PASS
wavekat-zh	zh_speech_mid.wav	0.1640	0.1640	0.0000	PASS

All rows within the existing 0.02 tolerance (max Δ = 0.0162).

Test plan

• make accuracy — 9/9 PASS
• cargo test -p wavekat-turn --no-default-features --features "" — clean
• cargo test -p wavekat-turn --no-default-features --features pipecat --test accuracy — 3 pipecat regression tests pass against the new schema
• cargo fmt --check, cargo clippy --features wavekat-smart-turn -- -D warnings — clean
• CI green

🤖 Generated with Claude Code