[AMD] Fix AITER unified decode page table for page size one

[vorapolsiloai]

Motivation

With AITER unified attention (SGLANG_USE_AITER_UNIFIED_ATTN=1), page size 1 (the default), and CUDA graphs enabled, decode reads a page table that was never filled in, so the output is wrong.

Before each CUDA graph replay, AiterAttnBackend._apply_cuda_graph_metadata() copies the request-to-token indices into the persistent cuda_graph_page_table. It did this for the sliding-window KV pool and for page_size > 1. It skipped page_size == 1, so the replay read the table's initial or stale contents. Eager decode (init_forward_metadata) builds this table for every page size, so only graph replay is affected. Query and KV dtypes do not matter.

I found this while validating #39998. The bug is in the existing decode metadata path, so the fix is a separate PR.

Modifications

• In the CUDA graph decode metadata path, always copy the page indices into cuda_graph_page_table. Convert them with _transform_table_1_to_real only when page_size > 1. This matches the eager path.
• Add test_cuda_graph_decode_copies_page_size_one_table. It calls _apply_cuda_graph_metadata() with page_size=1 and checks that the graph page table holds the req_to_token rows.

Accuracy Tests

• The new test fails without the fix (all 6 checked page-table entries differ) and passes with it on MI355X.
• test/registered/attention/test_aiter_fp8_q_unified_attention.py passes on MI355X.
• Re-validated post-rebase on AMD Instinct MI355X (ROCm 10, AITER): test_cuda_graph_decode_copies_page_size_one_table and all 4 unit tests in test/registered/attention/test_aiter_fp8_q_unified_attention.py pass cleanly.
• End-to-end check: Qwen3.8-27B MXFP4, MI355X, TP1, FP8 KV cache, AITER prefill and decode, CUDA graphs on. This stack also included [AMD] FP8 prefill for AITER: FP8-Q unified attention, and FP8 ASM FMHA for GQA ratios like 6 via KV-head repeat #39998.
  • GPQA: 48.48% (198 samples), the same score as the BF16-query reference (the stack without [AMD] FP8 prefill for AITER: FP8-Q unified attention, and FP8 ASM FMHA for GQA ratios like 6 via KV-head repeat #39998).
  • Needle-in-a-haystack: 100% at 64K and 128K (1,000 samples).
  • MMMU: 59.11% vs 60.78% for the same reference (900 samples). This is within our 2-point tolerance.

Speed Tests and Profiling

This is a correctness fix. It adds one device-to-device copy per graph replay, the same copy that page sizes above 1 already do. It is not meant to change performance.

Checklist

• Format code with pre-commit.
• Add unit tests.
• Provide accuracy results.

---

CI States

Latest PR Test (Base): ❌ Run #37836097801
Latest PR Test (Extra): ❌ Run #37836097439
Latest PR Test (AMD ROCm 10): ❌ Run #37836097712