add *.pyi stub generation support for bindings in rosetta.so

[htruscott]

Adds the --stubs flag to build.py for building pyrosetta, which uses pybind11-stubgen to inspect the C++ bindings generated by binder/pybind11 and generates *.pyi type annotation stub files in the rosetta directory. This allows static type checkers to see the members, arguments, and (most of the time*) parameter types included in pyrosetta. Adding **/*.pyi files to the setup.py package_data list means these stubs will be included automatically with any distribution, meaning it should be plug-and-play with a new release.

This PR adds a dependence on an external pypa package for stub generation (it's imported in the generate_stubs method so won't affect the process if unspecified) and I'm not familiar enough with the setuptools / ez_setup build pipeline to know how that should be integrated as a requirement. Thoughts? Should it be included as a submodule like other dependencies? Currently I just throw an error in generate_stubs in build.py if --stubs is specified but the package isn't included.

I think this will make the experience a lot easier for new users. I've tested it (installing from the .whl file in a fresh environment) in VSCode, and it should be compatible with static type analyzers and google colab as well!

*pybind11-stubgen uses the docstrings automatically generated by pybind11 to infer by-reference parameter types. If a module member is bound by pybind11 which references a type which has itself not yet been bound, the docstring can't use the python type name (since where the type should sit in the python module tree hasn't been specified) so it defaults to the C++ type instead (see the pybind11 documentation). Pybind11-stubgen will still include these C++ docstrings in the *.pyi files, which removes ambiguity for users most of the time [the user sometimes needs to connect the dots that std::shared_ptr<class core::pose::PDBInfo> corresponds to pyrosetta.rosetta.core.pose.PDBInfo] even if it can't properly annotate parameter types accordingly. This could theoretically be fixed by updating binder to run in two passes, binding all classes first and then all functions, but that would require a massive refactor for only marginal gain.