[310P] Size attention masks by actual context to avoid long-ctx OOM (v2, rebased on c460a15af)

[adeepn]

What

Re-derive the 310P long-context attention-mask OOM fix onto current upstream (base upstream-main @ c460a15af).

The base AscendAttentionMetadataBuilder.build() builds a dense attention mask sized to
max_model_len for every attention state. On 310P only PrefillNoCache consumes a dense mask; at
long context the O(max_model_len^2) allocation OOMs the NPU (observed: worker crash, "NPU out of
memory, tried to allocate 16 GiB", on a 131K-context model — the first inference dies in
get_attention_mask).

• Extract mask construction into a _build_attn_mask() hook on the base builder (default behaviour
  unchanged for all other devices).
• Override it in the 310P builder to return the dense mask only for PrefillNoCache, sized to the
  batch's actual max context via a new actual_max_seqlen arg. The size accounts for the alignment
  padding forward_prefill_310 folds into the last sequence (seq_len[-1] += num_actual_tokens - sum(seq_lens)), so the mask never under-covers the padded length.
• Build the SplitFuse (chunked-prefill) mask directly as [num_query_tokens, key_len] instead of
  materializing and index-selecting from a [max_seqlen, max_seqlen] tensor; grow the mask caches
  monotonically.

Why a new PR

Supersedes #22. Upstream rewrote the 310P attention/mask path (new compressed-mask / splitfuse-v2
code); #22 no longer applies. This re-derives the same fix onto c460a15af. Note: the upstream
compressed-mask path (_npu_flash_attention_v3) is unavailable on this CANN build, so the dense /
splitfuse paths are still taken and the fix remains required.

Test

• Unit (CPU, run in the 310P image): tests/ut/_310p/attention/test_attention_mask_310.py
  (tile rounding, direct splitfuse mask == index_select equivalence) and
  tests/ut/_310p/attention/test_attention_v1_310.py (mask only for PrefillNoCache; sizing covers
  the alignment-padded last sequence). 23 ut pass.
• Hardware: Qwen3.6-35B-A3B FP16, TP=4, 4×310P3, image nightly-main-310p (c460a15af): without
  this fix the first request OOMs in mask build; with it, loads, serves, KV budget ~1M tokens @131k,
  needle retrieval at short context, decode 4.4 t/s.