[mem_cache] Add free_full to release the full side of a tombstoned SWA node

[hnyls2002]

Summary

• Add free_full to the allocator family: a plain free() on the base, full-pool-only on the hybrid SWA allocators, a no-op on the all-SWA one
• Route the seven tombstoned-node releases through it, in both SWARadixCache and the unified tree
• Collapse the legacy tree's three evict-path releases into one _free_node_value helper, so the allocator call and the two eviction counters stay on the same branch

Background

• On a hybrid SWA allocator free() releases both halves: it gathers full_to_swa_index_mapping[idx], drops the entries that read as the padding slot, and returns the rest to the SWA pool
• A tombstoned node's SWA peers went back earlier (an SWA evict, or dec_swa_lock_only), and that cleared their mapping entries -- so free() on such a node already resolves to "full side only" today, purely because the sentinel filter absorbs it
• The caller knows which case it is. It just had no way to say so, and the filter is what stood between a mistake and the padding slot entering the SWA free list

Scope

• No behavior change: free_full(x) releases a subset of what free(x) releases, and on a tombstoned node the two are already equal
• What it buys is that a wrong caller becomes an assertable bug rather than silent SWA-pool corruption
• PureSWATokenToKVPoolAllocator gets an explicit no-op override -- its full_attn_allocator is its swa_attn_allocator, so inheriting the base default would free one pool twice
• All classes that define free_swa now define free_full as well; a hybrid subclass silently inheriting the base default is exactly the double free above

Also here

Free-group timing. _recover_tombstone_keeping_locked_full and the RecoverSWAWithLockedFull handler reached into alloc.full_attn_allocator.free(...) directly, which released immediately even inside a free group while every other path deferred to the barrier. They now go through free_full and defer like the rest.

Unified needs a distinct action type, not a flag read. FreeDeviceKV and friends are batched until a barrier, and the same batch's SWARebuild reassigns the node's full value before the batch drains -- by then the node's tombstone state no longer describes the value being freed. FreeDeviceKVFullOnly freezes the decision at emit time instead.

Test plan

• test_swa_unittest.py: TestFreeFullPartition -- the partition adds up to free, the SWA peers stay allocated, the mapping is left intact, the release defers inside a free group, and a single-pool allocator falls back to free
• test_unified_radix_cache_unittest.py: the RecoverSWAWithLockedFull case now asserts free_full instead of the inner allocator, so the free-group defer above is covered
• test/registered/unit/mem_cache passes

---

CI States

Latest PR Test (Base): 🚫 Run #33136294822
Latest PR Test (Extra): 🚫 Run #33136294751
Latest PR Test (AMD ROCm 7.2): 🚫 Run #33136294714