[HiCache] refactor: retire in-flight storage prefetches through one helper

[alphabetc1]

Motivation

Dropping an in-flight storage prefetch was written four times: release_aborted_request, the all-or-nothing discard in _check_hybrid_prefetch_result, revoke_pending_prefetch, and StorageAttachment._release_pending_storage_ops. Each releases the anchor host lock, the buffer-mode prefix context and anchor, the host pages, and the prefetch occupancy, and they had already drifted (two clamp prefetch_tokens_occupied at 0, two do not).

Modifications

• Add _retire_ongoing_prefetch(request, host_indices, extra_pools) and route all four through it. What each caller releases (the completed prefix, the aux pools, whether pool_transfers_done gates them) is passed in unchanged.
• _check_hybrid_prefetch_result loses four parameters it no longer needs.
• The occupancy decrement is clamped at 0 everywhere; this only differs from before when the counter would have gone negative.

Accuracy Tests

• test/registered/unit/{mem_cache,disaggregation,managers} and the Rust-core run of the shared suite: same results as main. The mock-based storage-cleanup test now wires the new helper like it already wires _prefetch_occupied_span.
• PD e2e smoke with decode HiCache + file L3 prefetch (Qwen3-8B), flushing L1/L2 between rounds: healthy (96 prefetches). That run mostly covers the success path; the discard paths are covered by the unit tests above.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.

---

CI States

Latest PR Test (Base): ⏳ Run #37202314137
Latest PR Test (Extra): ❌ Run #37202313945
Latest PR Test (AMD ROCm 10): ❌ Run #37202314347