refine: robustify computation of refinement regularization scale

[adelpit]

• Use the masked depth map for the distance calculation so it will consider only the content inside the user-defined region of interest
• Use the average of the inverse depth (harmonic mean of depth) so that the metric scales with parallax rather than depth. This will also have the benefit of increase the influence of near structures

[cdcseacave]

Hi Andrew, these days I had time to look over this PR, and I like the idea you are suggesting. However I was wondering if instead of computing a RegularizationScale for the entire image, would be better (in the same spirit of the idea) to compute it per pixel (or per triangle) like:

pixelRegularizationScale = SQUARE(depth / cameraA.GetFocalLength())

Hopefully this will be not only better, but cheaper to compute as we will not need to compute ComputeAverageLocalInverseDepth()

Did you try this already?