Phase 4: Radix-trie and block-pool unification (scheduler wiring)

[inureyes]

Context

This is the actual unification and the core of the epic. Today the radix trie stores whole dense DetachedCacheSet snapshots and adopt moves tensors into a new sequence, and the whole path is disabled under the paged backend. Make trie nodes reference physical blocks and reuse them across requests by refcount, reusing the copy-on-write primitive that already exists in cache/paged_detach.rs.

Tasks

• Change prompt-cache entries (src/server/prompt_cache/entry.rs, store.rs) to hold physical block references via DetachedPagedCacheSet / retain_block instead of dense DetachedCacheSet.
• Wire CachePool::detach_paged / adopt_paged into donate_finished_sequence_cache and try_adopt_cached_prefix (scheduler.rs).
• Remove the decode_storage_backend == Paged -> return None guard in try_adopt_cached_prefix and replace it with the paged adopt path.
• On a prefix hit, point the new sequence's block table at the matched physical blocks (refcount bump), with no tensor copy; the first write triggers the existing COW automatically.
• Reuse the APC block-hash chains as the block-granular match key so matching aligns to block boundaries.

Acceptance criteria

• Prefix reuse and paged batching run simultaneously (the documented "Dense backend only" limitation no longer applies).
• A hit reuses physical blocks with zero tensor copy, verified by observing refcount > 1 on the shared blocks.
• cached_tokens in the usage payload is reported correctly for paged hits.

Dependencies

Blocked by Phase 3 (paged prefill).

Part of #116