Replace homebrewed convolution/image filtering with OpenCV

This MR replaces the proprietary FFT convolvers with OpenCV's image filtering methods. This has several consequences:

1. Reproduciblity should be much better for peak finders because the boundary conditions for the convolvers was generally undefined.
2. Access to OpenCV's mature image processing algorithms. This should have performance implications in the long run.
3. Access to improvements in image processing. For example, we now use a simple binary thresholding, which can be replaced using e.g. Otsu's method, which is implemented in OpenCV.
4. The project is now architecturally much simpler. The convolvers added a huge amount of complexity and redundant functionality (box, delta convolvers never used).
5. Python scripting/testing is now much easier. Convolver classes introduced a lot of low level memory management to the python interface.

Resolves #474 (closed)