Replace all Mat creations with NumpyAllocator

[willprice]

Currently I'm pretty sure there's some nasty memory bugs waiting to surface and a lot of unnecessary data copying going on to account for this.

We should be taking advantage of the NumpyAllocator through the pyopencv_to and pyopencv_from functions in https://github.com/opencv/opencv/blob/master/modules/python/src2/cv2.cpp.

I'm not 100% sure how to do this, so need to do a spike first to figure out the details. A few places to draw inspiration:

• https://github.com/robotpy/robotpy-cscore/blob/master/src/ndarray_converter.cpp
• https://stackoverflow.com/questions/32481233/opencv-cython-bridge-leaking-memory
• https://gist.github.com/denis-trofimov/cd4c29dac34cf44d6c69f5802554802c]
• http://blog.enthought.com/python/numpy-arrays-with-pre-allocated-memory/
• https://codedump.io/share/cwWNW83GXltN/1/what-is-the-easiest-way-to-convert-ndarray-into-cvmat
• https://stackoverflow.com/questions/12957492/writing-python-bindings-for-c-code-that-use-opencv/12972689#12972689
• https://stackoverflow.com/questions/22736593/what-is-the-easiest-way-to-convert-ndarray-into-cvmat