Normalize CUDA Device to cuda:0

[FindHao]

fix #176

Changes

This PR adds device normalization logic to ensure all tensors are created on cuda:0, regardless of the device specified in the JSON configuration.

Modifications

1. load_tensor() function

• Added device normalization before loading tensor from file
• Any CUDA device string (e.g., cuda, cuda:1, cuda:2) is now mapped to cuda:0

2. _create_base_tensor() function

• Added device normalization before creating new tensors
• Applies to all tensor types: floating-point, integer, complex, and unsigned integer

Impact

• Prevents multi-GPU issues: All tensors are now consistently created on the same device
• Backward compatible: Non-CUDA devices (e.g., cpu) remain unchanged
• Minimal code change: Only two strategic locations modified for maximum coverage

Testing

This change affects the template file used for generating reproducers. All generated test cases will now place tensors on cuda:0 by default.

[meta-codesync]

@FindHao has imported this pull request. If you are a Meta employee, you can view this in D85061992.

[meta-codesync]

@FindHao merged this pull request in 197df27.