[ROCm] Add Torch SDPA and xFormers optimization for FireRedASR

[sammysun0711]

Hi FireRedTeam, thanks for your great work!

This PR aims to add FireRedASR optimization on ROCm on target platform AMD Instinct MI300+ GPU.

• Add docker/Dockerfile.rocm to quickly setup ROCm7 environment for deployment
• Add Pytorch SDPA and xFormer Attention support for performance optimization, can be controlled by environment variable: ATTENTION_BACKEND="SDPA" and ATTENTION_BACKEND="XFORMERS"
• Fix torch.load issue with weight_only=False for torch >= 2.6
• Add benchmark scripts and torch profiling support for performance analysis of different attention backend.
• Add FireRedASR optimization on ROCm guide in README.md.

Here are performance results with example audio (batch size=1) on single MI308X for your reference:

ATTENTION_BACKEND	RTF	Performance gain vs Native
Native	0.063	/
Torch SDPA	0.048	23.81%
xFormers Attention	0.056	11.11%

[kaituoxu]

Thanks for your PR, we will review.