Support Attention head_dim=512

[CC-Yeh]

• MatmulArguments: b_offset, b_leading_dimension, b_transpose (the last as a VARIANTS axis on gemm.metal).
• head_dim=512 dispatch (suffix > 8): per-group matmul(Q, K^T) → mask → softmax → matmul(P, V) → scatter. Smaller cases use existing single_pass/two_pass (also gained
  HEAD_DIM=512).
• New generic Softmax kernel under kernel/softmax/.
• Two attention helpers: ScatterScores, ScatterValues

[CC-Yeh]

Why BD=512 fused GEMM attention didn't work

• Threadgroup memory ceiling is 32 KB on Apple Silicon. Fused GEMM needs Q-tile + K/V-tile resident:

  BQ	BK	q_smem + kv_smem
  32	32	66 KB — overflows
  32	16	49 KB — overflows
  16	16	32 KB — at the limit
  16	8	24 KB — fits

• The only tile that fit (BQ=16, BK=8) was 2–3× slower than unfused matmul:

  seq	fused BD=512	unfused matmul	ratio
  512	4.23 ms	1.80 ms	2.4× slower
  2048	66.4 ms	22.5 ms	2.9× slower

• Why: BK=8 means 128 FMAs per KV reload vs SteelMatmul's 2048 (16× worse bandwidth amortization). BQ=16 means 2 simdgroup-matrix rows — too few to overlap math with loads.

Why we use the unfused matmul pipeline instead

• At HEAD_DIM=512 the work IS two big GEMMs (Q @ K^T and P @ V). GEMM kernels are already peak-tuned for those shapes; a hand-written attention kernel hardly beat them.
• MLX does the same — head_dim ∉ {64, 80, 128} → use_fallback = true → matmul + softmax + matmul at the graph level.