Pull out conv2d kernels from op

[lockshaw]

By pulling out the kernel functions into their own files, we should decrease build time by reducing the number of times the kernels (which rarely change) need to be recompiled

[reyna-abhyankar]

Overall, it looks good to me. I have one question that's related to the codebase. The .cpp file seem to mimic its corresponding .cu file. Is it auto-generated? If not, should that also be modified for this PR?

Also, what is the purpose of the kernel wrappers? If there's a forward and a backward and then also have forward_kernel and backward_kernel, I'm wondering if the wrapper is really necessary because it mostly seems to be doing some profiling.

[lockshaw]

Blocked on #345 as currently I can't check that the AMD kernel changes are correct

[reyna-abhyankar]

What is the purpose of the change from typename to class in model.h? Otherwise LGTM

[lockshaw]

What is the purpose of the change from typename to class in model.h? Otherwise LGTM

This is a C++17 feature apparently 🤷, whereas class is C++11 compliant. C++11 compliance checking is being added #271