feat(inference): opt-in fused Snake CUDA decoder

[zain1806481]

Summary

• Opt-in via KOKORO_DECODER_FUSION=1 after the model is on CUDA
• Freezes weight-norm modules and fuses Snake activations into one contiguous FP32 CUDA kernel (NVRTC)
• Fails closed when CUDA/NVRTC/FP32 requirements are not met; never enables a CPU fallback

Measured (Quadro K620, sm_50)

Alternating original/fused on the same frozen-weight model, fixed seed:

• ~8.5–8.8% wall-time reduction
• Exact waveform and duration match on every paired trial

Other GPUs may work (kernel targets the current device arch) but were not the validation target.

Test plan

• Boot API without the env flag (unchanged path)
• Boot with KOKORO_DECODER_FUSION=1 on CUDA FP32 and confirm log line Experimental decoder fusion enabled
• Compare a short synthesis against stock for audible/waveform parity on your GPU
• Confirm missing NVRTC fails closed with a clear error

Made with Cursor

[zain1806481]

Supporting notes for this PR:

• Measured decoder tuning (K620, exact audio): https://github.com/zain1806481/upstream-week-2026-09-14/blob/main/kokoro/DECODER-TUNING.md
• Upstream v0.9 review findings (voice cache / lazy phonemization priorities vs this opt-in fusion): https://github.com/zain1806481/upstream-week-2026-09-14/blob/main/kokoro/FINDINGS.md

Fusion remains opt-in via KOKORO_DECODER_FUSION=1 and fails closed without CUDA/NVRTC.