Update CUDA code to make it build on Windows

[vladkvit]

As per https://github.com/facebookresearch/pytorch3d/issues/843 - to solve the build issue on Windows, one must swap the header from <torch/extension.h> to <ATen/ATen.h>, and update torch:: references to at::. This replacement is the only change in this diff, and the only one that seems to be required.

Context - CUDA doesn't like pybind11, the torch/extension.h includes pybind11.

For reference, the error was "error: too few arguments for template template parameter "Tuple""

[ghost]

All CLA requirements met.