feat(moe): moe_mixture_forward kernel + Rust wrapper + unit test
Single-thread-per-(b,c) kernel, no atomicAdd, capture-friendly.
CPU-reference test verifies correctness for B=4, K=8, C=256 within 1e-6
tolerance.
Test wrapper uses mapped pinned memory exclusively
(feedback_no_htod_htoh_only_mapped_pinned.md): allocate MappedF32Buffer,
write CPU-side via host_ptr, kernel reads via dev_ptr, output via host_ptr
read after stream sync. NO memcpy_stod / memcpy_dtov anywhere.
GpuMoeHead struct in crates/ml/src/cuda_pipeline/gpu_moe_head.rs follows
existing cuda_pipeline head pattern (cubin loaded once, kernel handles
cached). Subsequent kernels (moe_mixture_backward,
moe_load_balance_loss, moe_expert_util_ema_update) extend the same
struct.
Spec: docs/superpowers/specs/2026-04-27-moe-regime-redesign-design.md §6.2.
Plan: docs/superpowers/plans/2026-04-27-moe-regime-redesign.md Task 2.1.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3d17c2a93c