[shihaobai]

GLM sparse MLA BMMs produce head-major outputs, requiring layout copies before prefill attention and when flattening multi-token V outputs. Write Q prefill and V prefill/decode results directly into token-major buffers through the existing bmm(out=...) interface, with explicit dimensions at each call site.

Validation on H200:

• 86 related GPU tests pass, covering BF16/FP16, 16/64 heads, padded noncontiguous inputs, guard rows, CUDA graph replay, and prefill/decode output projections.
• With real GLM-5.3-Flash layer-3 BF16 weights, decode V projection plus head flattening takes 16.61 → 12.90 µs at batch 32 and 22.75 → 15.68 µs at batch 128 (64 heads, five alternating AB/BA rounds). Batch 1 remains approximately 12 µs. These measurements cover projection and layout; full-model throughput was not measured.
• Repository-pinned Black, Flake8, and whitespace checks pass.