[DEV] Support indexed Mamba-2 inference on NVIDIA and MetaX

[big-hip]

Goal

Support indexed Mamba-2 SSD scan and single-token recurrence on NVIDIA and MetaX for InfiniLM model inference.

Scope

• Reuse the existing operator, workspace and graph plan/run interfaces; keep recurrent state in FP32 and validate indexed state ownership.
• Include the optional MetaX strict-FP32 GEMM policy used for Mamba continuation verification, preserving the default TF32 behavior.
• Keep graph allocator/recording fixes in shared runtime PR issue/1565 fix(runtime): support Qwen MTP and recurrent graph replay #1566.

Implementation and focused numerical/state tests are in existing PR #1562, paired with InfiniTensor/InfiniLM#575. This issue tracks the consolidated contribution; it does not request an additional feature PR.