Fix SWA eviction boundary and page-align chunked prefill

[hnyls2002]

Summary

Fix SWA radix tree insert boundary bug when page_size > sliding_window_size:

• _evict_swa could over-evict to exactly page_floor(seq_len), making all tokens being inserted fully evicted (case 3)
• _insert_helper had no handling for case 3 (swa_evicted_seqlen == total_length), falling through to _add_new_node(swa_tombstone=False) for already-evicted tokens
• This inflated swa_evictable_size_, caused negative usage, and led to potential double-free on subsequent tree eviction

Two-sided fix:

• Preventive: subtract extra page_size in _evict_swa formula so the frontier never reaches the insert boundary
• Defensive: add early return in _insert_helper for case 3 (free value, don't create node)

Test plan

• /rerun-test test_swa_eviction_boundary.py
• /rerun-test test_swa_unittest.py
• /rerun-test test_swa_radix_cache_kl.py

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

/rerun-test test_swa_eviction_boundary.py

[github-actions]

✅ 1-gpu-h100 (1 test): View workflow run

cd test/ && python3 registered/unit/mem_cache/test_swa_eviction_boundary.py

[hnyls2002]

/rerun-test test_swa_eviction_boundary.py

[github-actions]

✅ 1-gpu-h100 (1 test): View workflow run

cd test/ && python3 registered/unit/mem_cache/test_swa_eviction_boundary.py

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test test_swa_eviction_boundary.py test_swa_unittest.py test_swa_radix_cache_kl.py

[github-actions]

✅ 1-gpu-h100 (3 tests): View workflow run

cd test/ && python3 registered/unit/mem_cache/test_swa_eviction_boundary.py
cd test/ && python3 registered/unit/mem_cache/test_swa_unittest.py
cd test/ && python3 registered/radix_cache/test_swa_radix_cache_kl.py