Fix bug with hidden_to_memory and no hidden_cell

[drasmuss]

There was an error if hidden_to_memory=True and hidden_cell=None (since there is no hidden state to direct to the memory input). This handles that case. Note that this is technically equivalent to memory_to_memory=True, so it's kind of a weird thing to do, but I thought it was worth supporting so that people doing parameter sweeps and things like that can see the effect of this case rather than having to particularly avoid it.

[drasmuss]

After a bit more digging I was finding more and more awkward corner cases when you allow these hidden connections when hidden_cell=None. They're resolvable, but I think the "correct" behaviour is fairly unclear, and this is more likely to confuse users than be helpful (as well as making the code itself more complicated). So in the end I think it makes more sense to just not allow these cases (reversing what I said before). We don't lose any functionality (allowing hidden_to_memory is the same as memory_to_memory, as noted above, and input_to_hidden is equivalent to a skip connection, which can be created outside the LMU), and it makes everything cleaner.

So now this PR just adds a validation error for these cases.

[gsmalik]

I think everything looks great. I just added 3 very minor comments. Thanks!