[ROCm] Update and unify ROCm support for gfx942 and gfx950

[LZ-QWQ]

Summary

This PR updates and unifies slime's AMD support for ROCm 7.2 on gfx942 (MI300X/MI325X) and gfx950 (MI350X/MI355X) GPUs. It uses SGLang v0.5.15.post1, matching the SGLang version used by the current CUDA Docker image.

Changes

• Consolidate the ROCm image builds into a single architecture-selectable Dockerfile using ROCm 7.2 and SGLang v0.5.15.post1, and remove the obsolete variants.
• Refresh the AMD dependency patches and archive the previous patch set.
• Apply targeted fixes to the ROCm software stack for numerical consistency and runtime stability.
• Update the AMD Qwen3-4B launcher for the current slime workflow.
• Add English and Chinese AMD setup and troubleshooting guides.

Validation

Completed end-to-end Qwen3-4B validation on:

• 8× MI300X GPUs (gfx942)
• 4× MI355X GPUs (gfx950)

Follow-up work will expand ROCm validation and support to additional models, features, multi-node configurations, and automated CI coverage.