[AMD][MI355X] Bump MiniMax-M3 AgentX to latest ROCm nightly / [AMD][MI355X] 将 MiniMax-M3 AgentX 更新至最新 ROCm nightly (#2825)

* perf(mm3): bump MI355X AgentX ROCm nightly

Update only the MiniMax-M3 MI355X AgentX image pin to the latest immutable ROCm nightly, which includes the optimized BF16 indexer and routed GEMM.

中文：仅将 MiniMax-M3 MI355X AgentX 的镜像固定版本更新到最新的不可变 ROCm nightly，其中包含优化后的 BF16 索引器和 routed GEMM。

* chore: link MiniMax-M3 AgentX changelog

Replace the temporary changelog link with InferenceX PR #2825.

中文：将性能变更日志中的临时链接替换为 InferenceX PR #2825。

* fix(mm3): use full decode CUDA graphs

Keep ROCm breakable CUDA graphs disabled while selecting FULL_DECODE_ONLY explicitly, avoiding the new fail-closed piecewise-graph validation in the pinned vLLM image.

中文：保持 ROCm 可中断 CUDA Graph 关闭，并显式选择 FULL_DECODE_ONLY，以兼容固定 vLLM 镜像中新加入的分段 CUDA Graph 失败关闭校验。

Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>

---------

Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Chun Fang <chun.fang@amd.com>