musa: refine compute capability

[yeahdongcn]

Make sure to read the contributing guidelines before submitting a PR

This PR improves the handling of compute capabilities for MUSA devices with the following updates:

1. Adjusted Compute Capability Offset
   • GGML_CUDA_CC_OFFSET_MTHREADS is now positioned between NVIDIA and AMD.
2. Boundary Check Update
   • Updated the boundary check for NVIDIA compute capability tests using !GGML_CUDA_CC_IS_MTHREADS(cc).
3. Preserved Feature Availability Checks
   • Ensured that NVIDIA and AMD feature availability tests remain unchanged.

Testing Done

• ./build/bin/test-backend-ops
• ./build/bin/llama-cli -m ~/models/deepseek-r1_7b_q4_0.gguf -ngl 999

# ./build/bin/test-backend-ops
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 MUSA devices:
  Device 0: MTT S80, compute capability 2.1, VMM: yes
Testing 2 devices

Backend 1/2: MUSA0
  Device description: MTT S80
  Device memory: 16297 MB (16292 MB free)

  ABS(type=f16,ne_a=[128,2,2,2],v=0): �[1;32mOK�[0m
  ...
  4634/4634 tests passed
  Backend MUSA0: �[1;32mOK�[0m

Backend 2/2: CPU
  Skipping CPU backend
2/2 backends passed
�[1;32mOK�[0m

[yeahdongcn]

This is the initial PR. @fishingfly and I will collaborate to evaluate the features on MTT S80 and MTT S4000.

[yeahdongcn]

I re-ran the tests, and all passed.

[yeahdongcn]

Hi @JohannesGaessler Do you know how to retrigger the CI without pushing or force-pushing? Thanks.

[JohannesGaessler]

There's a button you can press as a collaborator.