remove redundant memory traffic

[FindHao]

In make_convolutional_layer function, l.output is initialized to 0 by xcalloc. Then it is copied to l.output_gpu, and conditionally copied to l.x_gpu and l.x_norm_gpu.

We can use cuda_memset to set these three arrays in gpu side rather than doing copy all zeros from cpu side. This optimization will save a lot of memory traffic. In my simple test for dog.jpg, it saves about 20% memory copy traffic.

[FindHao]

Between the initialization for l.output and the copies, l.output never changes.

[FindHao]

I found more examples with the same issue. And change cudamemset to async version.
For the simple test on data/dog.jpg, it gains 1.02x speedup on a RTX 2080Ti. This speedup is free lunch without harm in accuracy.