test(sdpa): fuzz head-axis stride gaps in the ragged sweeps

[vedaanta]

Follow-up to #516/#526. The with_ragged_token_gap knob fuzzes per-tensor TOKEN-stride gaps (nhd tokens). Head-axis gaps (stride[1] != d, e.g. head-interleaved buffers) are expressible by the native path in #526 (the TMA descriptors and Q/O offset math read stride[1] too) but are not fuzzed — extend the knob or add a sibling so the L0 ragged sweeps cover them, with gaps constrained to the 16-byte granularity rule.

🤖 Generated with Claude Code

[yuweih205]

Hi @vedaanta @Anerudhan, I'd like to help with #540. I can extend the Python L0 ragged SDPA sweep to cover head-axis stride gaps while preserving the 16-byte alignment constraints, and validate it on an H200/SM90 environment. Is this part still available, and are there preferred stride cases or test boundaries I should target? I'll keep the change focused and wait for your guidance before claiming the issue.