WGMMA+SP Kernel & AGU Serialization Fix

[NikhilRout]

1. Sparse WGMMA kernel

• old kernel redundantly loaded every warp’s compressed A and metadata through SMEM causing a lot of GMEM traffic
• now each warp only loads its own compressed A tile instead of cooperatively serializing loads across warps
• sp metadata is now directly loaded from GMEM -> TCU SRAM instead of old GMEM -> SMEM -> TCU SRAM
• now uses tilek/2 (A) tilek (B) strided descriptor layouts matching the TCU buffer

2. TCU metadata AGU serialization bug

• old AGU serialized all TCU blocks execution behind all warps' metadata memory round-trip latency. this prevented any memory-latency overlap
• changed to one in-flight TCU_LD transaction per TCU block with round-robin issue/commit and tagged response routing (just pipelining metadata load/per TCU)

---

Perf improvements

1.5x speedup on wg+sp+dxa
1.9x speedup on wg+sp