[AMD] Disable DSA dense decode graphs with DP attention

[taking-lying-flat]

Summary

• Do not enable the HIP DSA dense/sparse dual decode graphs when DP attention is active.
• Keep the existing dual-graph optimization enabled for HIP DSA deployments without DP attention.
• DP-attention deployments still use decode CUDA graphs; they capture the existing full-indexer graph, which is valid for all KV lengths.

Why

#31324 automatically enabled a dense k-only graph for DSA decode when max_kv_len <= index_topk. As isolated in #36071, replaying that graph on GLM-5.2 with DP attention can produce all-NaN logits on every attention-TP rank. The same topology is clean with eager decode and with the full-indexer graph, including when MTP and overlap scheduling are disabled.

This change applies the narrow correctness guard requested in #36071. It does not attempt to claim or mask a root-cause fix for the HIP k-only graph; that path should be re-enabled with DP attention only after its capture/replay buffer and stream contracts are qualified on the affected hardware.

Scope

• HIP + DSA + DP attention: use the correct-for-all full-indexer decode graph.
• HIP + DSA without DP attention: unchanged; retain dense/sparse dual graphs.
• CUDA and non-DSA models: unchanged.

Fixes #36071