[Fix] Forward SWA prealloc reclaim through the DSV4 HiSparse allocator

[mmangkad]

Summary

#39477 moved the decode-node SWA prealloc reclaim behind a new reclaim_for_prealloc allocator method, but DeepSeekV4HiSparseTokenToKVPoolAllocator wraps a SWA allocator rather than being one and hand-delegates each SWA method, so it never got one and every TP rank died with an AttributeError on the first admitted request. Delegate it to logical_attn_allocator and add a CPU regression case.

SGLANG_IS_IN_CI=true python test/registered/disaggregation/test_disaggregation_hisparse.py

Total latency: 15.536 s
Score: 0.960
Output throughput: 1199.603 token/s

Ran 2 tests in 154.649s

OK (skipped=1)

---

CI States

Latest PR Test (Base): ❌ Run #35445270465
Latest PR Test (Extra): ❌ Run #35445270148
Latest PR Test (AMD ROCm 10): ❌ Run #35445270461

[mmangkad]

/rerun-test test/registered/disaggregation/test_disaggregation_hisparse.py test/registered/unit/mem_cache/test_hisparse_allocator.py

[github-actions]

Results for /rerun-test test/registered/disaggregation/test_disaggregation_hisparse.py test/registered/unit/mem_cache/test_hisparse_allocator.py:

🚀 8-gpu-h200 (1 test): ✅ View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_hisparse.py

🚀 ubuntu-latest (1 test): ✅ View workflow run

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