migrate ETS to C++

[jmoralez]

Contributes to #753 by moving the functions that used numba in the ETS model to C++. Several changes had to be made to the library for this to work:

• Move the python package to a python folder. This is so that when running the tests from the root of the repository it uses the installed statsforecast package instead of the code from the statsforecast folder. With a regular python package these would be the same, however since we're now shipping a shared library these two are now different (the installed package has the shared library and the statsforecast folder doesn't).
• The tests now build the wheels first for the platforms that we'll support (windows amd64, macos x864_64, macos arm64, linux x86_64, linux aarch64 (arm64)) and then run the tests installing from that. These wheels will be published to s3 on pushes to main so that people can install from those without needing a compiler (the install notebook has instructions on how to do this).
• Updated the python-publish action to build the wheels for the same platforms as in the previous point.
• Added pybind11 to the dev requirements, which is used to build the shared library.
• Moved the restrict_to_bounds function and the results tuple from ETS to utils. These will be removed once we migrate all the models that use numba to C++.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[jmoralez]

@AzulGarza I think this is ready now. The changes since you last reviewed are:

• Switched to pybind11 to generate the python bindings. The main benefits are:

  • The overhead of calling a function is very low, which is great for objective functions that get called during numerical optimization.
  • The interchange between the two languages is easier, e.g. you can define that you want a double array and pybind11 will try to coerce the input to that if it isn't and raise an error otherwise, you can return tuples, etc.

  The downside is that we now have to build wheels for each python version instead of just one per os-architecture combination.

• Restored the setup.py file and extended it to build the shared library. This is so that we don't have to learn CMake, we just use that script.

• Updated the CONTRIBUTING.md file to use uv (since I deleted the conda environment files in fix ci #874 to have all dependencies in a single place) and added a section on how to rebuild the extension (I can extend on that if you want but it's just a matter of changing/creating the C++ files and rebuilding).

[AzulGarza]

this is amazing, @jmoralez! congrats🎉 lgtm