Added better rounding off of patching.

Added feature for rescaling to max input frame size.

Made face detection truly dynamic, with arbitrary image shapes.

First commit of single GPU inference