[Bugfix][Core] Accept needs_kv_cache_zeroing in SinkFullAttentionManager

[ivanium]

Purpose

Since #44490 the KV cache coordinator passes needs_kv_cache_zeroing to every single-type manager, but SinkFullAttentionManager still spelled out its parent's old __init__ signature. Any model using SinkFullAttentionSpec (openPangu with StaticSinkAttention) fails at KV cache manager init with TypeError: SinkFullAttentionManager.__init__() got an unexpected keyword argument 'needs_kv_cache_zeroing', whether or not zeroing is enabled.

This forwards **kwargs to the parent, matching the other manager subclasses (RSWA, SWA, chunked-local, Mamba, HiSparse), and adds a regression test that constructs the manager with needs_kv_cache_zeroing=True and checks its new blocks are recorded for zeroing.

Not a duplicate: no open PR touches this constructor (searched SinkFullAttentionManager and needs_kv_cache_zeroing; #47271 and #43374 modify nearby sink code but not this).

Test Plan

pytest tests/v1/core/test_single_type_kv_cache_manager.py tests/v1/test_kv_cache_spec_registry.py

Test Result

The new test fails on main with the TypeError above and passes with the fix; both files pass (81 passed). No model eval: the change only fixes a startup crash and does not affect outputs of models that already started.

AI assistance (Claude Code) was used; I reviewed every changed line and ran the tests above.

🤖 Generated with Claude Code

[ivanium]

This is not a urgent PR and I don't have a plan to merge it soon. Just wanted to leave it here so people and agents can refer to this if see any issues with SinkFullAttentionManager