[RFC] [bug] `logsumexp` in GCNNs

Created by: attila-i-szabo

At the moment, GCNNs fail if they only use real parameters but try to return a negative WF amplitude whose log should be complex. This arises if you use a nontrivial irrep, some entries of which are negative. The problem is logsumexp, which returns NaN if it has to take the log of a negative float. I am happy to fix this, but I would appreciate some input on which of the following strategies is preferable:

• Promote the outputs of the last layer into complex numbers – wastes memory
• Promote the characters into complex numbers – the code of logsumexp suggests that would waste the same amount of memory in type promotions
• Use the return_sign option of logsumexp and convert the sign part into 0 or iπ – this only works for real networks and irreps, so a lot of type checking and conditionals are needed

@chrisrothUT @PhilipVinc @femtobit?