Auto-generate Python bindings

[Krzmbrzl]

The current Python bindings are hand-written and they seem to be quite lacking in terms of what features of SeQuant are usable from Python.

Instead of using hand-written Python bindings, we could try to auto-generate them to make sure the Python side of things stays up-to-date with the C++ side in terms of features. We could even go as far as to auto-generate bindings in a dedicated Python submodule (e.g. sequant.auto) and then provide hand-written Python code in the main module that makes use of the auto-bound stuff to yield a nice and Pythonic Python interface. This could cover the most important bits and pieces while still allowing users a more raw access to a greater set of SeQuant features through the auto submodule.

Potential tools for auto-generating Python bindings:

• https://github.com/RosettaCommons/binder
• https://github.com/pthom/litgen

[evaleev]

I do not have much deep experience with these tools. Past experience with autogenerated bindings has not been perfect. Given how relatively young these tools are, adopting them would be a risk. I would be more inclined to first go back to the project I did try in the past and test https://github.com/wlav/cppyy which would also go one step further and avoid the need to maintain the bindings altogether.
P.S. in NWChemEx project cppyy was used in the beginning, but due to various technical reasons the python infrastructure got replaced by pybind. @ryanmrichard would remember the details better than I do and perhaps can have better pros/cons analysis..

[Krzmbrzl]

adopting them would be a risk

Not really - we could use the tools to generate the bindings and then just check them into git. To my understanding none of these tools are required at runtime 🤔

[ajay-mk]

Do we need to regenerate the Python bindings every time something changes on the C++ side?

[Krzmbrzl]

Well, if we add something that isn't exposed yet, we have to regenerate in order to make it available. But not doing so would still keep everything, that was available before, working.

With the Cling-based approach no binding generation would be needed (to my understanding). However, I'm not quite sure how user-friendly it is, if we require everyone to have Cling installed for the Python "bindings" to work. Let's wait for Ryan's insights on this approach.