Sphericart

- Minor bug fixes
- Runtime CUDA compilation via NVRTC
- CUDA wheels for `sphericart-torch` and `sphericart-jax`
- `metatensor` and `metatensor-torch` interfaces

- All APIs now offer spherical harmonics and solid harmonics (previously `normalized=True/False` respectively)
- New CUDA C++ and JAX APIs
- The Pytorch API now provides `SphericalHarmonics` and `SolidHarmonics` as subclasses of `torch.nn.Module`
- The Pytorch API warns (once) about second derivatives
- Solved bug when the spherical harmonics are called with zero points to calculate (all APIs)
- Various documentation improvements

/

julia-v0.0.3
This is the first public release of SpheriCart.jl, the Julia version of sphericart

Sphericart is a library for the efficient calculation of the spherical harmonics and their derivatives in Cartesian coordinates.

This release brings a lot of improvements:

- the CPU version of the code is now written in C++ instead of C, the main API entry point is the `SphericalHarmonics<T>` class
- the code is integrated with PyTorch and TorchScript in the `sphericart-torch` package, allowing backward propagation of gradients
- the `sphericart-torch` package also includes a CUDA version, computing spherical harmonics on GPUs
- the `sphericart-torch` package allows to patch e3nn, using sphericart implementation to compute spherical harmonics in e3nn-based models