changelog.txt

Changes with respect to e2cnn library:

######## escnn.group

- introduced GroupElements; group operations can be directly applied on instances of GroupElement (e.g. g * h or ~g)
  instead of using the methods from the group
- groups support multiple parametrizatiosn of their group elements
- groups support redundant subgroup ids which are then automatically parsed and converted in a unique id
- irreps identified by a tuple id rather than by their name. This id can be accessed and used to reconsrtuct the same
  irrep through `group.irrep(*id)`


######## escnn.kernels
- mostly refactored the package. Now use the WE theorem to automatically build the kernel basis instead of using the
  solutions manually found
- this package is now fully implemented in PyTorch, so its operations are differentiable and can be CUDA accelerated.


######## escnn.gspaces
- only two (actually 3) `GSpace`s (GSpace0D, GSpace1D, GSpace2D)
- each specific gspace is generated by a factory method and just instantiates one of the ones above
- for e2cnn's gspaces, same name and interface is used; since now these are methods and not classes,
  one needs to make the first letter lowercase (e.g. Rot2dOnR2() -> rot3dOnR2() )



######## escnn.nn
- R2Conv subclass of _RDConv
- basis expansion submodule moved outside conv so it can be used by Linear, too
- BasisExpansion abstract class becomes BasisManager interface (used also in escnn.nn.init)
- GeometricTensor supports optional argument `coords` (CAREFUL! each layer now has to manually check if the input tensor had coords and set them to the output too)
- FieldType.__call__() to generate GeometricTensors less verbosely
- FieldType.transform(tensor, el) supports also coords. If coords=None, only rotates coords.