Split equistore in a C-API core crate and a Rust wrapper

[Luthaf]

Closes #102
Closes #140

This is phase I of making equistore compatible with TorchScript, by having a separate C API we can load the code from a single shared library regardless if we are using it from Python/Torch/rascaline/…

• split equistore crate into a core, containing the C API and a Rust wrapper API
• the core no longer depends on the ndarray crate
  • loading files now takes a callback to create arrays in which the data will
    be placed
  • serialization is no longer optional, there is a single way the core can be
    built (it had multiple optional feature before)
• Python bindings explicitly manage the lifetime of array using Py_IncRef and
  Py_DecRef This is now Make Tensormap immutable #86
  • we no longer have to explicitly keep references to arrays around, this
    simplifies the code quite a bit

Still TODO:

• update the CI for the new configuration
• make sure all tests have been updated, and none is left commented out
• integrate with rascaline
  • check that the performance hit of not inlining equistore functions is not too bad
• write API documentation
• write something on the code organization for new developers
• remove all TODO in the code
• add soname to the shared library

---

📚 Documentation preview 📚: https://equistore--85.org.readthedocs.build/en/85/

[github-actions]

Here is a pre-built version of the code in this pull request: wheels.zip, you can install it locally by unzipping wheels.zip and using pip to install the file matching your system

[Luthaf]

And merged! 🚀