Adding ROCm Support

[SeanSong-amd]

This pull request enables model inference and training on AMD hardware with ROCm support. The key updates include:

1. Adjusted the CUDA kernel code in casual-conv1d and mamba-1p1p1 to be HIPify-friendly, utilizing PyTorch's HIPification routine to automatically generate .hip source files during package builds on ROCm.
   The original CUDA code remains intact to preserve existing functionality. All ROCm-specific modifications are wrapped within #ifdef/#ifndef USE_ROCM macros.
2. Minor Python Code Adjustments
   Made minor updates to the Python code to ensure compatibility with CUDA-based systems.

This push does not include ROCm-specific optimizations.

[SeanSong-amd]

@Unrealluver and the Vim team, Thank you so much for releasing the code! The community is truly excited to explore the new architecture. This PR introduces ROCm support, which will help make the model more accessible across a wider range of systems. I would be deeply grateful if you could kindly review the changes at your convenience and let me know if you have any questions or suggestions. Thank you!