feat(rocm/gemma4): integrate bc64 flash-attention prefill kernel (2.76x isolated) env-gated on gfx1201

Goal

Integrate the validated bc64 flash-attention prefill kernel (striped online softmax + WG256 packaging + PV 8-wave, isolated best 8744us = 2.76x over the current SharedK path at T=2048/d512, identity-verified) into the production ROCm gfx1201 prefill attention path, env-gated, default OFF.

Context

Isolated kernel development (thread on the bus) took the gfx1201 prefill attention from SharedK-class to 2.76x via three evidence-backed fixes: eliminating a 21x-redundant lane-replicated softmax, WG256 active-wave packaging, and PV 8-wave distribution. The remaining ~1.94x to Vulkan/ACO is dynamic-scheduling/codegen and was ruled a ceiling. This issue tracks landing the 2.76x construction into the serving path (projected ~1.63x whole-prefill once integrated).

Scope

• Add the bc64 kernel adjacent to PagedAttnPrefillSharedKWmma (src/vt/rocm/rocm_paged_attn.hip:1307), adapted to the paged-KV interface (block_table / seq_lens / query_start_loc), not the isolated contiguous form.
• Add a default-OFF VT_ATTN_PREFILL_BC64_FA host branch inside the eligible BF16/GQA block, before the SharedK dispatch (~:1694).
• Supported shapes only: d512 qg8 full window; d256 qg2 sliding window. Unsupported shape or env-absent/0 → byte-identical fallthrough.

Out of scope

• Replacing PagedAttnPrefillSharedKWmma. Importing the dirty lab VT_ATTN_PREFILL_* experiment matrix. Decode path. Non-BF16.

Gates

• off/default path byte-identical to current KEEP.
• on: shape reachability for d512-full + d256-sliding via real block_table/seq_lens/query_start_loc.
• token-exact vs the current SharedK path on real serving prompts BEFORE any perf number.
• same-binary A/B (flag on vs off) product p42k, fair PC=0, reverse-order medians.

Owner

Coordinator + hermes-vllm implement (don-agent offline, developer-directed); research fresh-review + p42k gate rerun.

Kind: feature

Start in src/vt/rocm/rocm_paged_attn.hip at PagedAttnPrefillSharedKWmma around line 1307, then inspect the host dispatch near line 1694 and the existing paged-KV inputs. Completion requires the default-off path to remain byte-identical, supported shapes to reach the bc64 kernel with real block_table, seq_lens, and query_start_loc data, and token-exact plus same-binary A/B validation to pass before performance comparison.