[None][fix] Fix VSWA gate in KVCacheManager window-size resolution

[eopXD]

@coderabbitai summary

Description

KVCacheManager.get_cache_indices, get_priority_by_block_id, and
get_batch_cache_indices gated the "window_size required" check on
len(max_attention_window_vec) > 1. That spuriously rejects a uniform
per-layer window vector such as [4096, 4096, ...], which has len > 1 but a
single effective window (is_vswa is False). commit_and_get_block_hashes
already used the correct is_vswa signal (distinct positive windows); the
others did not — so callers that legitimately omit window_size (e.g. the KV
cache connector) broke on any model with a uniform per-layer window vector.

This extracts a single _resolve_window_size helper keyed on is_vswa and
routes all four per-window accessors through it, removing the copy-pasted guard
and its divergence. Behavior for genuine VSWA (raise when window_size is
omitted) and for an explicitly-provided window_size is unchanged.

Test Coverage

tests/unittest/_torch/executor/test_resource_manager.py::TestResolveWindowSize

• test_uniform_multi_layer_vector_does_not_raise — the regression: a uniform
  [4096, 4096, 4096] vector resolves to 4096 instead of raising.
• test_genuine_vswa_requires_window_size — genuine VSWA still raises when
  window_size is omitted, and an explicit window_size passes through.

PR Checklist

• Please check this after reviewing the above items as appropriate for this PR.

[eopXD]

/bot run

[tensorrt-cicd]

PR_Github #58649 [ run ] triggered by Bot. Commit: 403dfdb Link to invocation