Resolves #547
Related to #578

This does a couple things:

1. Adds a workspace to Cuda that allows conv kernel to not have to re-allocate memory for patches
2. Removes a memset(0) from the output of Conv operation
3. Removes a memset(0) on patches allocation
4. No longer broadcasts filters in transpose_and_broadcast_filters
5. Sets patches[i] = 0.0 for both unfold_input & unfold_output
6. Uses the parallel stream to parallelize conv operations a bit

Timings of conv2d bench on A10