feat: add macOS MPS device support to all PyTorch models

[Ayush10]

Summary

Closes #2050

Add automatic MPS (Metal Performance Shaders) device support for Apple Silicon Macs across all PyTorch model benchmarks.

Approach

Introduced a centralized get_torch_device() utility in pytorch_utils.py that selects the best available device with priority: CUDA > MPS > CPU.

def get_torch_device(GPU=0):
    if isinstance(GPU, str):
        return torch.device(GPU)
    if torch.cuda.is_available() and GPU >= 0:
        return torch.device(f"cuda:{GPU}")
    if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")

Updated all 27 model classes and 2 module-level device selections to use this function, replacing the previous inline pattern:

# Before (in every model file)
self.device = torch.device("cuda:%d" % GPU if torch.cuda.is_available() and GPU >= 0 else "cpu")

# After
self.device = get_torch_device(GPU)

Files changed

• qlib/contrib/model/pytorch_utils.py — Added get_torch_device() alongside existing count_parameters()
• 25 model files in qlib/contrib/model/ — Updated device selection
• qlib/contrib/model/pytorch_tra.py — Updated module-level device
• qlib/contrib/data/dataset.py — Updated module-level device

Backward compatibility

• CUDA users: No change in behavior. CUDA is still preferred when available.
• CPU users: No change. CPU is the final fallback.
• String GPU parameter: Still supported (e.g. GPU="cuda:1").
• The hasattr(torch.backends, "mps") guard ensures compatibility with older PyTorch versions that lack MPS support.

Test Plan

• All 27 model device selections updated
• count_parameters() function preserved unchanged
• hasattr guard for PyTorch versions without MPS backend
• String GPU parameter handling preserved (pytorch_nn.py special case)