[mem_cache] Make free_swa sync-free on page_size == 1

[hnyls2002]

Summary

• free_swa on page_size == 1 gathers the peers and releases them without the swa_indices > 0 filter; the filter's data-dependent output shape read a count back to the host on every SWA release
• The filter's first job, dropping padding entries, is no longer needed: callers declare intent with free_full for released peers ([mem_cache] Add free_kv_row to release a request's kv row by row range #36721), grouped frees resolve peers at enqueue ([misc] Resolve SWA ownership at enqueue time for grouped free() #36646), and duplicate insert slices are split at the eviction floor ([mem_cache] Split duplicate insert frees at the SWA eviction floor #37481), so the gather never reads a padding entry
• Its other job, absorbing a caller that frees a range whose peers are already gone, becomes a contract: swa.peer_mapped, an async Invariant (Bucket.FATAL_UNCONTAINABLE) checked at the gather under SGLANG_INVARIANT_CHECK
• The mirror contract guards free_full: swa.peer_released (mapping[free_index] == 0, Bucket.GUARD) fires when a caller releases only the full side of a slot whose SWA peer is still live, which would otherwise leak that peer; free_group_end frees the pending full ids directly so the check runs once, at enqueue
• page_size > 1 keeps one filter per group in _release_swa: HiCache LOAD_BACK re-pairs a page-aligned full chunk with an offset SWA one, so a page can hold unmapped slots

Tests

• TestSWAPeerMappedContract: the checked predicate is false exactly for a slot whose peer is gone, and free_swa no longer trips torch.cuda.set_sync_debug_mode("error") (CUDA only, gated on the pre-fix form still being flagged)
• Drop test_free_group_owns_tombstoned_indices: free() on a tombstoned slot is the caller error the contract now reports
• SWA fixtures that declare an eviction floor now release the peers below it first (free_swa) before calling free_full / inserting with swa_evicted_seqlen, as window eviction does; test_free_full_leaves_the_mapping_intact is dropped because the mapping is always zero at that point under the contract
• test_self_e2e_perturb_req_to_token runs with SGLANG_INVARIANT_CHECK=0: a perturbed slot reaching free_swa with its peer released is the corruption under test, not a regression

Follows #36646.

---

CI States

Latest PR Test (Base): 🚫 Run #33678830613
Latest PR Test (Extra): 🚫 Run #33678830328
Latest PR Test (AMD ROCm 7.2): ⏳ Run #33678830257

[hnyls2002]

Latest PR Test (Base): ❌ Run #33573528101
Latest PR Test (Extra): ✅ Run #33573527976

PR chains passed