Context

Neuron Python bindings have been implemented by using the Python C API directly. Among the major issues we identified:

• ref-counting, which is simply too easy to get wrong, especially for routines that can exit early.
• Identifying which references should be borrowed or stollen, which requires a careful read of the API and reference management.
• The API verbosity. Operations like looping over a Python list are verbose, error prone and not particularly intuitive, incurring a high maintenance toll.

Such problems lead to "undead" objects i.e., objects which are no longer in use but due to its wrong ref count they are not garbage collected, keep using valuable memory.

This PR

In this PR we bring nanobind to improve our lives and apply it to solve a case where there was such leaking of undead objects.

TODO

• Find a way to avoid git update --recursive as this is only needed for nanobind
• Try to expose symbols on windows without using dllexport or do it in a separate header file.