bf16 medium DiT: FMHA-fused TRT engine build recipe + CLI (bf16 = medium default, fp16-mixed kept)

[Cortexelus]

bf16 medium DiT — FMHA-fused TRT engine

Adds a bf16 DiT precision for medium as the new speed default, and wires --precision through the CLI and SA3Inference. fp16-mixed (canonical) and fp32 remain fully selectable.

What changed

• Build recipe (build/build_from_onnx.py, build/build.py): new sa3-m-bf16 target. Reuses the same raw fp32 sa3-m/dit.onnx — no new ONNX export — and builds with BuilderFlag.BF16 + EXPLICIT_BATCH (not STRONGLY_TYPED). bf16 carries fp32's dynamic range, so the FP32-softmax islands that fp16-mixed had to keep vanish and TRT 10.15's FMHA fuser fires (0 → 96 fused _gemm_mha_v2 nodes). Build-time precision recipe only — no graph change.
• Precision registry + CLI (scripts/sa3_trt_core.py, scripts/sa3_trt.py): PRECISIONS gains bf16; per-model default resolution via default_precision(). --precision now defaults to None → resolves to bf16 for medium, fp16mixed for sm-music/sm-sfx. bf16 is medium-only — sm-music/sm-sfx use standard attention and already fuse in fp16-mixed, so requesting bf16 for them errors out.
• README: --precision section with the per-model default table.

Speed (H100 SXM, medium)

~1.77× @ L=256 and ~4.70× @ L=4096 vs fp16-mixed, within the perceptual re-seed floor (FAD 0.59× floor, CLAP within spread, n=128).

Reproducibility caveat

bf16 is not seed-reproducible vs fp16-mixed. The medium DiT uses differential attention (cancellation-sensitive), so bf16 is a different-but-equal take per seed — quality within the re-seed floor, but exact samples differ. Use --precision fp16mixed when you need bit-reproducibility or max per-step fidelity.

Engine distribution

The dit_bf16.trt engine is gitignored and ships via HF (stabilityai/stable-audio-3-optimized), auto-downloaded on first use. No engine binary is included in this PR.

Notes

• The TRT DiT engines are static batch=1 (the ONNX bakes batch=1), so CFG runs as a sequential cond+uncond dual-pass at batch=1 for every precision — unchanged by bf16.
• Separate / orthogonal bug: a pre-existing medium + same-l + cfg>1 malloc crash exists and is not related to this change (it also reproduces on fp16-mixed). Tracked separately.

Branched off bbd30b5.