Add ROCm / AMD Instinct MI300x support

[subhajitdchow]

Summary

Adds AMD Instinct MI300x support via ROCm 7.2, PyTorch 2.9.1, FlashInfer v0.5.3+amd.2, and CK-backend flash-attn (pinned to 0f82fea).

NVIDIA path is unchanged — all new logic is gated on torch.version.hip.

What's included

• setup_rocm.sh — one-command ROCm setup inside FlashInfer Docker container
• ssd/layers/flash_attn_compat.py — SDPA fallback when native flash-attn unavailable
• Dual attention backend (sgl_kernel on NVIDIA, flash_attn/SDPA on AMD)
• Dual tree-decode backend via SSD_TREE_DECODE_BACKEND={flashinfer,sdpa}
• Runtime-conditional FlashInfer plan() args and default CUDA arch
• Qwen3 rope_scaling fix, return_dict=False for newer transformers

Validation

70B target + 1B draft, 5× MI300x: ~225.86 tok/s (4.32× AR, 1.64× SD) across alpaca / c4 / ultrafeedback / humaneval.