[mem_cache] Split duplicate insert frees at the SWA eviction floor

[hnyls2002]

Summary

• _insert_walk_step splits the duplicate slice of an insert at the request's swa_evicted_seqlen: FreeDeviceKVFullOnly below it, FreeDeviceKV above it, in both the Python and the Rust tree core
• Matches what SWARadixCache._insert_helper already does through free_kv_row_segments

Why

• Below the floor the slice's SWA peers were already released by window eviction, so a both-side free there releases them a second time
• Today free_swa reads mapping = 0 for those slots and the swa_indices > 0 filter drops them, at the cost of one host sync per release; once that filter becomes a contract ([mem_cache] Make free_swa sync-free on page_size == 1 #36723) the same free trips swa.peer_mapped
• Both insert producers reach this one emission site: the regular cache_unfinished_req insert and the HiCache load-back insert (floor = span_end - len(swa_dev))

Tests

• test_dup_slice_below_eviction_floor_frees_full_only asserts the emitted actions in the shared unified suite, so it runs under both tree-core backends; fails on the current code with no FreeDeviceKVFullOnly emitted
• The eight components/swa.rs overlap tests whose duplicate slice lies entirely below the floor now expect FreeDeviceKVFullOnly
• Full test_unified_radix_cache_unittest.py and cargo test --locked pass

---

CI States

Latest PR Test (Base): 🚫 Run #33573510228
Latest PR Test (Extra): ✅ Run #33573509920
Latest PR Test (AMD ROCm 7.2): 🚫 Run #33573510276

[hnyls2002]

/tag-and-rerun-ci

[Jialin]

Could you help to confirm if we have async assert inside token_to_kv_pool_allocator.free_full to confirm if there's no swa mapping from the full slots? I feel that could better guardrail the change to avoid leakage.

sglang/python/sglang/srt/mem_cache/unified_radix_cache.py

Line 1084 in 1aa8299

elif isinstance(action, FreeDeviceKVFullOnly):

[hnyls2002]

For this PR, the changes should be no-ops. The whole stack is to remove [swa_indices > 0], I added an async assert here, https://github.com/sgl-project/sglang/pull/36723/changes#diff-5e4f4819b718cd7f28b375e1d1729d525290e1f8fa79da7353ccad7c2ad3b56cR368

Which has real behavior changes.

[Jialin]

Thanks for your clarification. Yeah, ideally, this change should be no-op.

However, I think that assertion you mentioned verify existence of SWA peer in free_swa, while FreeDeviceKVFullOnly should expect no SWA peer (which should be served as a long term invariant of FreeDeviceKVFullOnly usage. And if we use async assert, the overhead should be small enough). WDYT?

[hnyls2002]

Yes, I will add an extra guard of expect(swa.peer_released, mapping[free_index] == 0) in #36723