Support free-threading Python wheel distributions

As some of you might be aware, since CPython 3.13, there is experimental support for an alternative interpreter distribution (via [PEP703](https://peps.python.org/pep-0703/)) that disables the GIL altogether, thus enabling support for real thread parallelism.

Since PyTorch now supports free-threaded wheel distributions (https://github.com/pytorch/pytorch/issues/130249), it should be easy producing wheels for fairscale as well.

In order to provide free-threaded wheels, there is a need for extensive testing, which can be summarized in performing the following tasks:

- [ ] Audit Python bindings and declare them free-threading compatible (xref https://py-free-threading.github.io/porting/#updating-extension-modules).
- [ ] Run the test suite with pytest-run-parallel to find potential issues, and fix them.
- [ ] Run the test suite under ThreadSanitizer. Should be feasible, since there are no runtime dependencies aside from Cython.
- [ ] Add cp313t-* to CI to build free-threading wheels.

I can volunteer in taking charge of these tasks if necessary

Start with the Python bindings and the free-threading porting guidance linked in the issue, then run the test suite with pytest-run-parallel. Investigate ThreadSanitizer results and the existing CI wheel configuration. Done means the bindings and tests are compatible with free-threading and CI builds cp313t-* wheels.

Written by the indexing model from the issue text.