Optimized box blur implementation

[RunDevelopment]

A box blur is a separable filter, so I used OpenCV's sepFilter2D to implement box blur. However, I noticed that the DFT implementation of filter2D eventually catches up to sepFilter2D and ends up being faster for large radii. So I only use sepFilter2D for radii below 70. That's the point of which both are about equal performance-wise.

The speed-up is especially noticeable for small radii. E.g. a radius of 10 on a 2k RGB image takes around 0.08 sec with sepFilter2D and around 0.27 sec with filter2D.

Also, just to clarify: this is a pure optimization. The outputs of filter2D and sepFilter2D are the same.