feat: add AMD ROCm support via AITER CK flash attention backend

[ZJLi2013]

Summary

Add AMD Instinct MI300X GPU support to Matrix-Game-3.0 using AITER (AI Tensor Engine for ROCm) Composable Kernel (CK) flash attention backend.

• attention.py: Add AITER CK dispatch path between FA3 and FA2. Dispatch priority: FA3 → AITER CK → FA2 → RuntimeError. No changes to existing NVIDIA code paths.
• README.md: Add AMD ROCm installation section with ROCm 7.x + AITER setup instructions.

Motivation

The original code hard-asserts flash_attn availability, which fails on AMD GPUs. This PR adds native AMD support via AITER's CK backend, which compiles flash attention kernels to native AMD ISA (hipcc → LLVM → GCN).

Test Results (MI300X, gfx942, ROCm 7.2)

Backend	Iter 2 (steady-state)	Step time	vs FA2 Triton
FA2 Triton	14s	~5.0s	baseline
AITER CK	11s	~4.0s	-25%

• Tested with: 1× MI300X, 2 iterations, 3 inference steps, 704×1280, seed=42
• Output: identical 1.8MB MP4 video across all backends
• Environment: rocm/pytorch:rocm7.2.1_ubuntu22.04_py3.10_pytorch_release_2.9.1 + amd-aiter 0.1.13

Changes

File	Lines	Description
Matrix-Game-3/wan/modules/attention.py	+32/-4	AITER CK import, dispatch path, error handling
Matrix-Game-3/README.md	+19/-2	AMD MI300X support note, ROCm install section

Compatibility

• NVIDIA: Zero impact — existing FA3/FA2 paths unchanged
• AMD ROCm 7.x: AITER CK backend (recommended, best performance)
• AMD ROCm 6.x: FA2 Triton backend still works via flash-attn with FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
• No AITER or flash-attn: Clear RuntimeError with install instructions (replaces silent AssertionError)