bf16+tiled (m=128) GEMMs at M=4096 — 1072 GOPS, default serving kernels

bf16 MiniLM GEMMs (M=512..8192, qkvfused) — NPU beats CPU 2.2-3.2x

Optimized MiniLM GEMMs: M=512..8192 + QKV-fused (NPU beats CPU 1.37x)

4-col MiniLM-shape GEMMs (384/1536 @ M=512), verified in real model forward

Add MiniLM model-shape GEMMs (384/1536), verified in real model forward

Initial: proven Phoenix NPU1 IRON kernels + fusion experiment