Encode runs on Vulkan, decode stays on XNNPACK. Exported with the
conv1d-as-conv2d lowering, the slice-based STFT magnitude, hoisted
cross-attention K/V and a head-major KV layout. The int8 build quantizes
every decode-side linear; encode is identical between the two.
Requires the ExecuTorch 1.4.1 Vulkan runtime from the v0.10.0-libs-1.4.1
artifacts.
Add int8 MLX variant; remove bf16 (OOM-killed on device)