Add INT8 VAE decode engine

[ryanontheinside]

Summary

• Adds an INT8-quantized TRT engine for the VAE decoder via TensorRT entropy calibration. 1.64x faster than fp16 on standalone 60s decode (33ms vs 54ms), 32.5 dB PSNR vs fp16, +88 MB VRAM vs fp16 (TRT retains kernels for both precisions).
• Worth shipping for workloads where standalone VAE decode dominates (large vae_window, batch generation). For tight streaming with short vae_window and skip-cache reuse, the speedup dilutes to ~1.7% end-to-end because the diffusion decoder dominates the tick. See docs/int8_vae_decode.md for the full decision guide.
• Full research narrative including hybrid + autotune negative results lives on branch archive/ryanontheinside/int8-research-2026-05-01.

What's in this PR

• acestep/engine/trt/vae_decode_int8.py: INT8 builder with entropy/minmax calibrators and optional first/last conv pinning
• scripts/build_vae_int8_engine.py: canonical CLI driver
• scripts/collect_vae_calibration.py: regenerate calibration latents (32 prompt-diverse, 1500-frame each)
• tests/benchmarks/bench_vae_int8_regular.py: regression bench harness (synthetic latent + wav round-trip)
• demos/test_stream_cover_graph.py: adds --vae-decode-engine and --source-audio flags so streaming runs can A/B engines
• docs/int8_vae_decode.md: full variant comparison report (matrix, streaming bench, VRAM analysis, decision guide, what-didn't-work)

Test plan

• Generate calibration latents: uv run python scripts/collect_vae_calibration.py --frames 1500 --output-dir ~/.daydream-scope/models/demon/calibration/vae_latents_60s
• Build engine: uv run python scripts/build_vae_int8_engine.py --engine-name vae_decode_int8_60s --opt-frames 1500 --max-frames 1500
• Run bench: uv run python tests/benchmarks/bench_vae_int8_regular.py --int8-engine vae_decode_int8_60s --fp16-engine vae_decode_fp16_60s and confirm >=1.5x speedup, >=32 dB fp16-vs-int8 PSNR
• Streaming sanity: uv run python demos/test_stream_cover_graph.py --vae-decode-engine vae_decode_int8_60s --vae-window 5 and confirm output is audibly comparable to the fp16 baseline run
• A/B listen to bench_outputs/vae_int8/vae_decode_int8_60s/synth_trt_int8.wav vs synth_trt_fp16.wav (synthetic latent reveals the decoder difference; wav round-trip masks it because the encoder dominates)