This is not only not used anywhere, but could also try to initiate an RV with incompatible shapes and size

* Allow signature to handle rng and size arguments explicitly.
* Parse ndim_supp and ndims_params from class signature
* Move rv_op method to the SymbolicRandomVariable class and get rid of dummy inputs logic (it was needed in previous versions of PyTensor)
* Fix errors in automatic signature of CustomDist
* Allow dispatch methods without filtering of inputs for SymbolicRandomVariable distributions

This allows sampling from multiple backends without having to dispatch for each one