yet more cross-talk

I figured out that the fastest and most robust estimate of the cross-talk would come from not the entire chip but from a high-variance small sub-image of the cross-talking chip, and that the fit I do should not be to all pixels in the image section, but only to pixels that are simultaneously extreme-valued in the cross-talking chip and non-extreme-valued in the cross-listening chip. This is now what the code fits. Tomorrow the job is to figure out which chip-chip pairs have significantly non-zero cross-talk amplitudes and modify mosaic_mrdfits() so that it can make use of that output.