[compute_numeric_gradient] Use torch.flatten instead of view(-1).

[mhaz]

I ran into technical issues when doing the convolutional network exercise from assignment 3 for the 2019 iteration of the class.
The issue was the same as #4 and stemmed from using .view(-1) to flatten the f(x+h) and f(x-h) tensors in the numerical gradient evaluation. Using flatten instead seems to fix the problem.

Unfortunately, I ran into another runtime error the "Batchnorm for deep convolutional networks" subsection. It seems to stem from the implementation of FastConv, but this might be more than I can chew right now.