[DSv4] Fix shared index cache accessor page size

[kpham-sgl]

DSv4's non-paged FP8 indexer prefill crashes because DeepSeekV4IndexerPool does not expose the storage page size expected by the shared DSA accessor. #42178 introduced this dependency; #42671 renamed the field to index_page_size. Reported in this comment.

Set index_page_size to the pool's physical page size.

Validation:

• Pre-commit and git diff --check passed.
• Manual H200 check using real DSv4 pools and the unmodified Triton gather: main raises the missing index_page_size error for 64- and 128-slot pages; the PR returns exactly the expected keys and scales for two sequences with non-sequential pages and partial final pages.
• Manual DSv4 Flash FP8 server comparison on 4 H200s, TP4: a 9000-token prefill enters the non-paged gather on all four ranks with its default enablement and 8192-row threshold. Main crashes with the reported attribute error; the PR completes prefill and generates 8 tokens (HTTP 200).
• Server comparison used a 16384-token prefill chunk, 32768-token cache budget, SWA/full ratio 1, and disabled graphs/radix caching to keep the full request eligible. DeepGEMM compiled on demand. This was a functional reproduction, not an accuracy or performance benchmark.

---

CI States

Latest PR Test (Base): ❌ Run #37577779907
Latest PR Test (Extra): ❌ Run #37577779610
Latest PR Test (AMD ROCm 10): ❌ Run #37577779910

[kpham-sgl]

/rerun-test test/registered/unit/mem_cache/test_dsv4_compressed_pools.py test/registered/disaggregation/test_disaggregation_dsv4.py

[github-actions]

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

cd test/ && python3 registered/unit/mem_cache/test_dsv4_compressed_pools.py

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

cd test/ && python3 registered/disaggregation/test_disaggregation_dsv4.py

[kpham-sgl]

Manually verified on 4×H200 (TP4) with DSv4 Flash FP8: a 9000-token prefill hits the non-paged path with its default 8192-row threshold. Main crashes with the missing index_page_size; this fix completes generation successfully. Real GPU gather checks also pass for 64- and 128-slot pages. Test changes have been reverted; the PR is now a one-line fix.

[kpham-sgl]

/rerun-test test/registered/disaggregation/test_disaggregation_dsv4.py

[github-actions]

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

cd test/ && python3 registered/disaggregation/test_disaggregation_dsv4.py