This PR is an attempt to fix #2146, although not in the way that I would have hoped. Unfortunately, it seems that for now, for the sake of memory safety, we will need to copy the input matrix.

Specifically, the problem is this: layers like Convolution and AtrousConvolution and TransposedConvolution (and some others) create an alias of the given input matrix when Forward() is called; this is the member inputTemp. This alias uses the same memory as the given input matrix; it does not make a copy, and it is not responsible for freeing the memory when the alias's destructor is called. That member inputTemp may then be used in Backward() and Gradient().

However, the problem is that there exist layer types that will cause the input matrix to other layers to be freed during the forward pass; this means that any layer that took an alias during the forward pass is now holding a pointer to invalid memory. (I can't remember if it is individual layers that cause the freeing, or if that happens as part of the FFN and RNN classes. In either case, the result is the same.)

I thought that perhaps we could simply use std::move() instead of creating an alias, but this turns out to not work because other parts of the FFN and RNN forward/backward passes rely on having the layer input still available even after Forward() is called on the layer. So, for now, I've simply copied the input. That's not the fastest solution, but I think it's better to have a slow solution that works instead of a fast one that sometimes segfaults. :)

Some points for discussion in this PR:

• If the layer input is still expected even after Forward() is called by the FFN class, then should forward really take arma::mat&&s for both input and output? Or should the input be, e.g., const arma::mat&?

• Have we written anywhere the specific expectations with what can and can't be done with the inputs to Forward() implementations? (I guess that's kind of related to the first question.)

• Is it possible to refactor FFN and RNN such that the std::move() approach (instead of the copy approach I used here) can work?

I ask these questions because I'm not particularly familiar with the internal design of this part of the code, so maybe there is just something I'm not aware of, I don't know. :) Relevant folks for this discussion: @zoq, @ShikharJ, @walragatver, @sreenikSS, @saksham189. (Sorry for so many tags. It just seems to me like there are a couple of fundamental issues here that might be worth discussing. 👍)