Stats16b again

Trying to fix again #700 (closed)

Problem arises when statistics have been computed only for some of the images in a first pass of normalization/stacking. The next time around, the remaining images stats are computed in [0,1] while image stats already computed are converted back to [0,65535].

=> Strategy is to write stats in seq (and in cache) always consistent the bitdepth of the original image.

This should be done with fbfe24e9. Next step is to simplify statistics_internal_ushort now that only valid statistics are passed (no need to convert back anymore). And to test extensively.

EDIT: Conversion factors removed in statistics_internal_ushort, normalization and sequence_export. Testing done in 8, 16 and 32b, should be ok now...