[Perf] Hoist DSv4 draft-extend SWA write locs; unify SWA graph buffer naming

[hnyls2002]

Hoist the DSv4 draft-extend SWA write-loc translate out of the captured graph, and align the remaining backends on the shared buffer conventions.

DSv4 draft-extend

• Previously the out-graph metadata init left swa_out_cache_loc unset, so every SWA layer's store_cache recorded a full_to_swa_index_mapping[out_cache_loc] translate inside the graph (one per layer per replay).
• init_cuda_graph_state now pre-allocates cuda_graph_swa_out_cache_loc (draft runner only); the metadata init refills it outside the graph (zeroed tail, [:num_tokens] view) and stamps it, so the captured store_cache takes the cached path -- one translate per step, off the replay.
• Eager miss (no graph state / oversized batch) stamps a fresh per-step translate instead of falling back to per-layer translation at store time; capture always fits the pre-sized buffer, so it never records a temporary tensor's address.

Cross-backend alignment

• Rename swa_out_cache_loc_buf -> cuda_graph_swa_out_cache_loc in the flashattention and ascend backends; the SWA write-target graph buffer now has one name across triton / trtllm_mha / flashinfer / aiter / fa3 / ascend / dsv4.
• XPU: read full_to_swa_index_mapping live from the pool instead of caching the tensor ref at init; drop a redundant pool re-assignment.

---

CI States

Latest PR Test (Base): 🚫 Run #32104845985
Latest PR Test (Extra): ❌ Run #32104845896

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test test_deepseek_v4_flash_fp4_b200.py test_deepseek_v4_flash_fp8_h200.py test_swa_radix_cache_kl.py

[github-actions]

Results for /rerun-test test_deepseek_v4_flash_fp4_b200.py test_deepseek_v4_flash_fp8_h200.py test_swa_radix_cache_kl.py:

🚀 4-gpu-b200 (1 test): ✅ View workflow run

cd test/ && python3 registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py

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

cd test/ && python3 registered/models_e2e/test_deepseek_v4_flash_fp8_h200.py

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

cd test/ && python3 registered/radix_cache/test_swa_radix_cache_kl.py