E3nn-jax

Added
- `e3nn.reduced_symmetric_tensor_product_basis(irreps: Irreps, order: int)`
- `e3nn.IrrepsArray.filtered(keep: List[Irrep])`
- `e3nn.reduced_tensor_product_basis(formula_or_irreps_list: Union[str, List[e3nn.Irreps]], ...)`
- `IrrepsArray.at[i].set(v)` and `IrrepsArray.at[i].add(v)`
- add `Irreps.is_scalar`

Added
- Simple irreps indexing of `IrrepsArray`: like `x[..., "10x0e"]` but not `x[..., "0e + 1e"]`
- `e3nn.concatenate, e3nn.mean, e3nn.sum`
- `e3nn.norm` for `IrrepsArray`
- `e3nn.tensor_product`
- `e3nn.normal`
- Better support of `+ - * /` operators for `IrrepsArray`
- Add **new operator** `e3nn.grad`: it takes an `IrrepsArray -> IrrepsArray` function and returns a `IrrepsArray -> IrrepsArray` function
- Add support of operator `IrrepsArray ** scalar`
- Add support of `x[..., 3:6]` for `IrrepsArray`
- Add `e3nn.reduced_tensor_product_basis`
- Add `e3nn.stack`

Removed
- `IrrepsArray.cat` is now deprecated and replaced by `e3nn.concatenate`
- `e3nn.full_tensor_product` is now deprecated and replaced by `e3nn.tensor_product`
- `e3nn.FullyConnectedTensorProduct` is now deprecated in favor of `e3nn.tensor_product` and `e3nn.Linear`
- **breaking change** remove `IrrepsArray.from_any`
- **breaking change** remove option `optimize_einsums`, (it is now always `True`)

Changed
- **breaking change** rewrite the `equivariance_error` and `assert_equivariant` functions

Changed
- **breaking change** change the ordering of `Irrep`. Now it matches with `Irrep.iterator`.
- **breaking change** `Irrep("1e") == "1e"` and `Irreps("1e + 2e") == "1e + 2e"` are now `True`.
- **breaking change** `Linear` simplify the `irreps_out` which might cause reshape of the parameters.
- `index_add` supports `IrrepArray`

Added
- broadcast for `Linear`
- argument `channel_out` to `Linear` for convenience
- `Irreps` can be created from a `MulIrrep`
- `"0e" + Irreps("1e")` is now supported
- `"0e" + Irrep("1e")` is now supported
- `map_back` argument to `index_add`
- `IrrepsArray.split(list of irreps)`
- `poly_envelope` function

Changed
- **breaking change** rename `IrrepsData` into `IrrepsArray`
- **breaking change** `IrrepsArray.shape` is now equal to `contiguous.shape` (instead of `contiguous.shape[:-1]`)
- **breaking change** `IrrepsArray * array` requires `array.shape[-1]` to be 1 or `array` to be a scalar
- **breaking change** `IrrepsArray.contiguous` is renamed in `IrrepsArray.array`
- **breaking change** `IrrepsArray.new` is renamed in `IrrepsArray.from_any`
- `spherical_harmonics` normalization is now set to `component` like everything else.

Removed
- **breaking change** `IrrepsArray.from_contiguous` is removed. Use `IrrepsArray(irreps, array)` instead.

Added
- add `e3nn.config` to set global defaults parameters
- `__getindex__` to `IrrepsData`
- `gradient_normalization` argument that can be `element` or `path`
- `path_normalization` can be a number between 0 and 1
- add nearest interpolation for `zoom`, default is linear
- implement `custom_jvp` for spherical harmonics

Added
- Docker image

Added
- add the `sh` function that does not use `IrrepsData` as input/output
- `legendre` algorithm to compute spherical harmonics
- add flag `algorithm` to specify the algorithm to use for computing spherical harmonics, use `legendre` for large L.
- `experimental.voxel_convolution`: add optional dynamic steps (not static for jit)