Fix rnn+symmetry training

[limymy]

fix #121

[ClemensSchwarke]

Hey @limymy, thanks again for the PR :) I don't think we can support recurrent training with symmetry this way, because we would need to be more careful with the hidden states of the network. Simply repeating them should not work because those would not be the hidden states the network produces when encountering the mirrored states "naturally". Does that make sense?

[limymy]

You're right! In a time series, the movement of different legs will lead to different hidden states, so they cannot be simply copied. Do you have a better idea to solve this problem?

[ClemensSchwarke]

I think it is not trivial, I will add an error message for now. If you have a good way to implement this let us know! Otherwise, we will adress this later as we currently work on other parts of the library.

[ClemensSchwarke]

Hey @limymy, I added the error message in #125. I would close this PR for now until we have a proper way of handling the hidden states. Thanks again for bringing this up and creating the PR, it might become useful later :)