Change default initializer for DenseSymm

Created by: PhilipVinc

This is one piece of the discussion in #863 : initializer of DenseSymm.

This changes the initializer to have correct default variance of 1 for every channel. The implementation is complicated because the kernel itself has a shape that is not related to the size of the input n_sites which is what should be used as variance scaling.

(Also includes #870, so ignore the first commit)

@chrisrothUT please review and approve if you agree.