Enable MiniMax-M3 vLLM plugin path

[lirui927]

Motivation

Enable MiniMax-M3 vLLM plugin path.

Technical Details

• Add MiniMax-M3 sparse MHA support in the vLLM plugin path.
• Align sparse MHA KV cache handling with the page-shuffled layout expected by ATOM kernels.
• Reuse vLLM-provided output buffers in sparse prefill/decode to avoid extra allocation and copy.
• Handle actual-token slicing for padded vLLM batches and zero padded output tails.
• Route mixed decode+prefill sparse batches through the prefill path so per-token sparse block tables are built correctly.

Test Plan

• Start vLLM service and run full GSM8K.

Test Result

atom native

vllm-atom

## Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.