Reset a sub-image when it is diverging

Divergence is detected when then end peak of a major iteration is a factor of a few (4 by default) higher than the start peak. If other sub-images are not finished yet, they will ask for another major iteration that could again result in divergence. This should be ok for a first implementation -- it may produce an acceptable image.

This MR was tested inside Rapthor, and it indeed avoided divergence: a valid image was created. The results were not optimal because during the divergence one of the sub-images hit the niter limit (which typically happens during a diverging deconvolution). As a result, wsclean recovered the subimage but did stop afterwards because of the niter limit. In a next MR I will also check inside GenericClean and MultiScale to make sure to stop early when diverging, so that niter is not hit.