[Bugfix][HiSparse] Adopt GPU prefix copies after the hit's allocation

[MatthewBonanni]

Purpose

Fix an engine crash (Cannot get N free blocks from the pool) on a HiSparse host-prefix hit under GPU memory pressure. This is an alternative to #57457.

A host-prefix hit re-pins free GPU copies of its pages so the prefix comes back GPU-resident. Admission counts those copies as free blocks, but adoption pinned them before the hit's own allocation (hot region and new pages), which could then run the pool dry.

#57457 fixes this by charging the pins at admission. That mixes host and GPU accounting in HiSparseSourceManager.get_num_blocks_to_allocate, and under pressure it defers a request that could run by reading its prefix from host, just to keep an optimization.

Changes

• Adopt after allocation. commit_computed_blocks records the pages to adopt, and update_residency (in cache_blocks, after allocate_new_blocks) adopts the copies that survived. The allocation always gets what admission counted, so no admission accounting is needed. A copy the allocation reuses is dropped by its reuse watcher and never adopted.
• Copies are evicted last. BlockPool.unpin_blocks documents unpinned blocks as last-resort eviction candidates, but blocks freed afterwards were appended behind them, so an allocation evicted a prefix's copies before blocks nobody would reuse. FreeKVCacheBlockQueue now keeps a last-resort segment at its tail that later appends stay in front of, and unpin_blocks uses it.
• Tail first. HiSparse releases a prefix's pages tail first, matching free_blocks' convention, so pressure trims a prefix from its end.

Replaying #57457's scenario (a resumed 3-page host prefix hit, free_blocks left in the pool):

free_blocks	main	#57457	this PR
5–6	rejected	rejected	rejected
7	crash	rejected	admitted, 0/3 copies
8	crash	rejected	admitted, 1/3
9	crash	rejected	admitted, 2/3
10+	3/3	3/3	3/3

Test Plan

• test_hisparse_prefix_hit_under_pressure_adopts_surviving_copies[7, 9, 10]: exact adopted pages, and no block shared between adopted copies and fresh allocations. 7 fails without the adoption timing, 9 without tail-first release, and 10 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:

python cpu_pytest.py --confcutdir=tests/v1/core tests/v1/core/test_prefix_caching.py tests/v1/core/test_kv_cache_utils.py

Test Result

• New tests fail on main (3 of 4) and pass with this PR; each case was also checked against the partial fixes above.
• 297 passed. 8 errors are fixture 'caplog_vllm' not found, from running outside the top-level conftest, and are identical on main.

No GPU or model run: the change is scheduler-side block bookkeeping.

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

🤖 Generated with Claude Code

[MatthewBonanni]

Superseded by #59282 (stacked with the optional #59283), which carries this fix from a vllm-project branch so the follow-up can stack on it.