[Bugfix] Fix full-graph FIA seq length alignment in TND layout

[LICO1314]

What this PR does / why we need it?

• Align actual_seq_lengths_q with padded query token count when running full-graph FIA updates to satisfy TND layout requirements.
• Move the alignment helper into vllm_ascend/attention/utils.py so future FIA-based update paths can reuse it.
• This prevents queryT != actualSequenceLengthQ[-1] mismatches that lead to aclnnFusedInferAttentionScoreV3 tiling failures under full-graph mode.
• Fixes # (optional: link CI failure or issue if you have one)

Does this PR introduce any user-facing change?

No. This is a bug fix in full-graph parameter update alignment and does not change public APIs.

How was this patch tested?

Not run locally. Reproduced/observed in CI (full-graph + speculative decode) and the fix is a targeted alignment change.

• vLLM version: v0.14.1
• vLLM main: vllm-project/vllm@dc917cc

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request addresses a bug in the full-graph FIA path by ensuring actual_seq_lengths_q is correctly aligned with the padded query token count. The introduction of the align_actual_seq_lengths_q helper function is a good, reusable solution. My review focuses on the correctness and safety of the changes. I've identified a critical issue in full_graph_fia where an IndexError could occur with empty request batches, and I have provided a recommendation to fix it.

[gemini-code-assist]

The logic to determine num_tokens across lines 534-536 is not safe. attn_metadata.actual_seq_lengths_q can be an empty list if there are no requests in the batch, which will cause an IndexError when [-1] is accessed on line 534.

This logic can be rewritten to be both safe and more concise. Consider replacing lines 534-536 with the following:

        unpadded_num_tokens = attn_metadata.actual_seq_lengths_q[-1] if attn_metadata.actual_seq_lengths_q else 0
        num_tokens = max(unpadded_num_tokens, int(query.shape[0]))