Replace legacy HGEMM kernels with gfx950 A16W16 universe GEMM

[xytpai]

Motivation

The existing HGEMM implementation was fragmented across separate split-K and small-M kernels, creating duplicated logic and maintenance overhead. This PR consolidates them into a single gfx950 A16W16 universe GEMM supporting FP16/BF16, all tensor layouts, split-K, slice-K, small-M, and HTI paths, with unified tests and benchmarks.

Supported Features

• gfx950 target
• FP16 and BF16 inputs
• Input-dtype or FP32 output
• Optional bias
• NN, NT, TN, and TT layouts
• Dynamic shapes and leading strides
• Padded strides and storage offsets
• Configurable M/N/K tile sizes
• Configurable pipeline stages
• Configurable M/N/K wave decomposition
• Workgroup-local slice-K via k_waves
• Cross-workgroup split-K with atomic reduction
• K-tail and split-partition tail handling
• Small-M and boundary-tile handling
• Half-tile interleaved scheduling
• Direct global-to-LDS asynchronous loads
• Swizzled and transposed LDS layouts
• Vectorized C-shuffle and global stores
• Packed FP16/BF16 atomic accumulation
• XCD-aware block swizzle
• Configurable group_m
• Layout-dynamic compilation cache
• Correctness, padded-stride, and benchmark coverage