Fixes conv transpose stride bug, adds more docs to upscale2d

[coreylowman]

Resolves #657

There were two bugs here:

1. patches stride was incorrect in the actual kernel - was using op.chan_in instead of op.chan_out
2. In the backward pass gemm_batched call, it was not broadcasted the filters (so using m * n instead of 0)

Additionally I added a bunch of tests and cleaned up the implementations.

I also changed upscale methods to accept the method as a paramter. This is for a number of reasons:

1. This allows us to actually add args to each of the methods in the future, and use them in the methods.
2. It unifies the compile & run time shape calls.

[coreylowman]

The interface to upscale2d has changed. It now takes in the method as a parameter so you can call it like in these examples (notably without specifying any generics in either case). This unifies how you call them in both const/runtime situations, as before you would've had to specify the method in the generic both ways