Add an engine backend: the 22-block DiT runs on hand-written WebGPU
kernels (tiled/sgmat GEMM, fused LN+adaLN, RoPE, tiled QK/softmax/AV,
grouped convpos, GRN) with 4-bit packed weights and per-group scales
dequantized in-kernel. The worker loads only the ORT vocoder for this
path; ONNX Runtime remains the fallback.
- 229.7 MB int4 bundle, ~1.2 s per clip on WebGPU (was 2.8 s int4 ORT)
- flow: 4 steps, sway -1, cfg 0 (B=1, no null pass)
- tuning: tts.backend / tts.gemm switch kernels without touching the worker
- fixes found on the way: u32 vs f32 uniforms, second SiLU on the time MLP,
RoPE inv_freq exponent, AV score-tile column, f32 attention scores,
safe tanh in Mish
