[ROCm] Tighten persistent sparse-MLA disable predicate

[ChuanLi1101]

Small hardening on top of vllm-project#47567 (mitigation for vllm-project#47042).

• Slice step_query_lens / total_seq_lens to num_reqs and cast both to int64 so the chunked-prefill-continuation mask cannot broadcast-mismatch under cudagraph padding.
• Clarify in the comment that decode is single-token for this backend (UNIFORM_SINGLE_TOKEN_DECODE, no spec-as-decode), so gating on query_len > 1 does not touch the decode fast path (addresses the ''don't gate decode'' concern).

Behavior is unchanged for the current shapes; this is defensive correctness + documentation. Merge this into the vllm-project#47567 branch to include it there.

Co-authored-by: Claude

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[Rohan138]

@ChuanLi1101 my PR has already been merged - if the int64 cast is specifically needed (not sure why, cudagraph padding shouldn't change dtypes? All our tests are passing rn) please open a new PR to vLLM upstream.