(AKA Issue with grain merge layer mode)
Fall back to using the legacy grain-extract/merge modes when the
input image precision is integer, and only use the non-legacy modes
when the precision is floating point. The midpoint intensity value
of the non-legacy modes is 0.5, which isn't representable exactly
using integer precision, causing small deviations between the input
image and the reconstructed image.
Add an option to keep the decomposition in a layer group.
Add an option to add layer mask to each scales layers.
Do not use 'new from visible' because it produces unexpected result when
- image is already composed by several layers
- target layer has not the same size as the image canvas
Replaced by succession of layer copy and merge down.
