Phase 5: Block-budget admission, eviction, and preemption

[inureyes]

Context

With physical blocks shared across requests, scheduling and eviction must be tied to the global block budget rather than per-sequence buffers. Freed prefix blocks must return to the pool when no sequence references them.

Tasks

• Admit or queue sequences based on free block count; integrate with the existing PreemptionPolicy.
• LRU/TTL eviction over radix-tree leaves that releases physical blocks when their refcount reaches 0; merge this with the current prompt-cache capacity/TTL accounting (prompt_cache/policy.rs, store.rs).
• Implement recompute-on-preempt (drop blocks, re-prefill on resume). Host swap-out is out of scope because unified memory makes it low value.
• Surface pool block usage in GET /v1/cache/stats and the Prometheus gauges (extend prompt_cache/metrics.rs).

Acceptance criteria

• Under block pressure the scheduler evicts cold prefixes and admits new work without OOM.
• Metrics report block-level usage (allocated / live / free blocks and bytes).

Dependencies

Blocked by Phase 4 (radix-trie / block-pool unification).

Part of #116