sdpa_vector_2pass_1_gqa kernel batch offset for K/V in gqa decode

[dudududukim]

• ✅ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: Claude was used as coding agent and i've cross-checked the written codes.

---

Fix batching idx for sdpa_vector_2pass_1_gqa kernel's K/V pointer

Following my recent PRs(#4077, #4380) introduced sdqa_vector_2pass_1_gqa for eliminating redundancy reading for each KV head, there was bug when using batched decode without mask.

So the batch idx is added for pointer update and properly works for batch inputs.

kernel	batch offset	status
sdpa_vector (single-pass)	folded into tid.x → kv_head_idx = q_batch_head_idx / gqa	correct
sdpa_vector_2pass_1 (non-GQA)	explicit kv_batch_head_idx = batch_idx*num_kv_heads+kv_head_idx	correct
sdpa_vector_2pass_1_gqa	batch term missing	THIS FIX

Test for before fix and after fix (following CONTRIBUTION.md)

• New test case (B, L) = (2, 8192) added to test_sdpa_vector_gqa_long: FAILS on upstream main (AssertionError), PASSES with the fix.
• No regression: B == 1 is byte-identical (batch_idx == 0 → kv_batch_head_idx == kv_head_idx), so all existing (B==1) benchmark shapes are unchanged by construction. bench_sdpa_gqa.py two-build compare confirms within measurement noise (the gqa-4 control, identical code on both builds, drifts as much as the gated kernels — that drift is the noise floor).
• pre-commit (clang-format / black / isort) passes on the 2 files.

Formatting

uvx pre-commit run --all-files ✅ PASSED

[zcbenz]

Thanks!