build: get torch from PyPI except as dev dependency on Windows (#717)

### Summary of Changes

* For MacOS, get `torch` from PyPI. This fixes an installation error.
* For Linux, get `torch` from PyPI. It already includes CUDA support.
* For Windows, get `torch` from
`https://download.pytorch.org/whl/cu121`, but only as a dev dependency.
Anyone using the library, always gets the default `torch` version
without CUDA from PyPI, unless they explicitly change that. The source
defined in `pyproject.toml` is completely ignored.

Author: lars-reimann