[feat](minimax): support minimax-2.5 in atom-vllm mode

[PerryZhang01]

Motivation

This PR adds support for MiniMax-M2.5 1 in ATOM vLLM plugin mode, following the onboarding flow in add-atom-vllm-model.md. Initial bring-up surfaced gaps between the native ATOM server path and the plugin path that required extra wiring and fixes.

• Fused QKV in plugin mode: The vLLM plugin attention stack expects a fused qkv buffer not only separate q, k, v tensors.
• Config propagation: Some HF / runtime flags (e.g. trust_remote_code) are applied correctly at the vLLM CLI layer but were not always forwarded into ATOM’s plugin Config.
• Kernel / layout edge cases: Remaining none bugs tied to paged-attention assembly kernels.
• server script need add --trust-remote-code param.

so, currenly we use eager mode to run minimax-2.5 in atom-vllm mode, and these problems will be done in next prs.

Test Result

TODO

• remove qkv in atom-vllm plugin mode and use q,k,v, so qkv cat can be removed
• add eps in reshape_and_cache_with_pertoken_quant ops to avoid div zero, so plugin can use graph mode
• fix none bugs in paged-attention assembly kernels if avaiable.