[Core][HiSparse] Keep unpinned blocks last in the free queue

[MatthewBonanni]

Purpose

Optional performance improvement, stacked on #59282. #59282 fixes the crash on its own; this PR only affects how many HiSparse GPU prefix copies survive memory pressure, and so how many prefix pages are read from host instead.

BlockPool.unpin_blocks documents unpinned blocks as "last-resort eviction candidates", but it appends them to the free queue like any other block, so blocks freed afterwards are queued behind them. In #59282's scenario, a finished request's cached indexer block is freed after its MLA copies were unpinned, so the resumed request's allocation evicts a copy it was about to adopt while that indexer block survives.

Changes

• FreeKVCacheBlockQueue keeps a last-resort segment at its tail: append/append_n insert in front of it, the new append_last_resort_n appends behind it, and popleft/popleft_n/remove advance its head when they take it.
• BlockPool.unpin_blocks uses append_last_resort_n, which makes its documented behavior true.

Only the HiSparse coordinator unpins blocks. Without unpinned blocks the segment is empty and the queue orders blocks exactly as before, at the cost of one pointer comparison per popped or removed block.

This does change HiSparse's eviction policy: unpinned blocks, meaning finished requests' copies and pages a running request released but still reads, now outlast every ordinary cached block rather than aging in LRU order. Evicting either kind forces reads from host.

Copies kept in #59282's scenario:

free_blocks	#59282	this PR
7–8	0/3	0–1/3
9	1/3	2/3
10	2/3	3/3
11+	3/3	3/3

Test Plan

• test_hisparse_prefix_hit_under_pressure_adopts_surviving_copies[10] now expects all three copies; it keeps two without the last-resort segment.
• test_free_kv_cache_block_queue_last_resort_stays_behind: later appends stay ahead of the segment after its first block is removed or popped.
• tests/v1/core/test_prefix_caching.py and tests/v1/core/test_kv_cache_utils.py, with CpuPlatform selected.

Test Result

• 296 passed. 8 errors are fixture 'caplog_vllm' not found, from running outside the top-level conftest, and are identical on main.
• Not yet measured: an end-to-end pressure benchmark showing fewer host reads. That should decide whether this is worth a core change.

AI assistance (Claude Code) was used for the investigation, implementation, tests, and this description.

🤖 Generated with Claude Code