perf: stateful streaming VAE decode — eliminate redundant overlap

[KevinAHM]

Summary

• Streaming decode re-decoded 4 overlapping patches through the VAE each step, discarding 75% of the output
• Replace with stateful decode that carries causal conv padding buffers between calls — one patch in, one patch out, no overlap
• StreamingVAEDecoder caches CausalConv1d and CausalTransposeConv1d left-pad state, matching the approach used in the WebGPU ONNX VAE port

Changes

• Add StreamingVAEDecoder to audiovae/audio_vae_v2.py
• AudioVAE.streaming_decode() context manager for clean lifecycle
• _inference yields single-patch latents in streaming mode (was 4-patch chunks)
• _generate and _generate_with_prompt_cache use StreamingVAEDecoder

Benchmarks

Streaming VAE decode (isolated, 35 patches):

Method	Time	Max diff vs full decode
4-patch overlap (before)	289ms	0.0011
Stateful (after)	148ms	0.0005

2x faster VAE decode, and more accurate (half the error of the overlap approach).

Correctness

• Cosine similarity generate() vs generate_streaming(): 1.0000
• Tested with voice cloning (reference audio + timbre transfer)

Test plan

• Cosine similarity: stateful decode vs full decode
• A/B audio comparison with voice clone (Nellie reference)
• End-to-end generate() vs generate_streaming() match
• Streaming output length matches non-streaming

🤖 Generated with Claude Code