Summary

Fix GPSampler crash when torch.set_default_device("cuda") is set globally.

When users set torch.set_default_device("cuda"), GPSampler would crash because torch tensors created internally would be placed on CUDA while numpy arrays remain on CPU, causing conversion errors like:

TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

Changes

• Wrapped the sample_relative computation in a torch.device("cpu") context manager to force all tensor operations to use CPU
• Extracted the main computation into _sample_relative_impl for cleaner separation
• Added regression test that verifies GPSampler works with CUDA default device (skipped if CUDA not available)

Test plan

• Added regression test test_gpsampler_with_cuda_default_device
• Existing tests should pass

Fixes #6113

cc @kAIto47802 (who suggested this approach in the issue comments)