Account for batch size

[andrewfurey21]

Currently mlpack layers return gradients that are the summed across a batch, when typically they should be averaged otherwise gradients will change drastically if batch size changes.

[rcurtin]

I agree with this change---basically we'd need to go through all differentiable optimizers and divide the step size by the batch size. This would be a "breaking" change so I would want to do a major version bump. We could also get the Bandicoot, FP16, and maybe FASTA/FISTA/FBS PRs merged before that too. I'll see if I can find the time to open a PR for it soon.

[rcurtin]

I went through all the separable optimizers and opened #431 with this fix. It actually turned out to be far less work than I thought!

[rcurtin]

Fixed since #431 is merged.