Add tests: Slice assign vs Cat in LSTM backward

[louisfd]

In order to fix #1098, where it is mentioned that LSTM backward does not work in its current state but would if slice_assign were replaced by cat in its computations, I wrote some tests: one for slice_assign backward, and one specifically for LSTM backward, both comparing the result to the one we get with cat. For both tests it worked fine, making me unsure if there is really a bug with slice_assign.