MUSA: Use Monkey Patching to Automatically Convert CUDA Backend to MUSA

[yeahdongcn]

This PR introduces monkey patches to automatically convert CUDA backend to MUSA backend.

Key Updates

1. util.torch_auto_backend.py
   • Added global variables CUDA and CUDA0 to replace hardcoded "cuda" and "cuda:0".
   • Implemented monkey patching to automatically convert CUDA backend to MUSA.
   • Added test cases for MUSA.

Testing Done

• make dev_install
• python ./tests/torch_auto_backend_test.py

  Torch backend loaded: CUDA=musa, CUDA0=musa:0
  musa musa:0
  tensor([1.2000, 2.3000], device='musa:0')
  tensor([1.2000, 2.3000], device='musa:0')
  True
  8
  <torch_musa.core.device.Device object at 0x7ff626246320>
  tensor([1.2000, 2.3000], device='musa:0')
  0

• Ran inference using (the output tokens still have some problems):

  numactl -N 1 -m 1 python ./ktransformers/local_chat.py --cpu_infer 33 \
    --model_path deepseek-ai/DeepSeek-R1 \
    --gguf_path /models/hub/models--unsloth--DeepSeek-R1-GGUF/snapshots/02bcc0a0f68146dae57942804d82bdf0cc636003/DeepSeek-R1-Q4_K_M \
    --optimize_rule_path /ws/ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml

[Atream]

prefill_device and generate_device is not needed to change.
Change it in your custom yaml file. It's the same elsewhere.

[yeahdongcn]

I just want to ensure consistency throughout the code (Both CUDA and "cuda" will be present.).
Would you like me to revert this change?

[Atream]

Thank you for your contribution.
We will merge after planning and testing a unified architecture that is compatible with various GPUs. Until then, please use your own branches first. Remember to frequently merge the main branch to stay synchronized with us and achieve better performance.