Add GDK_MEMORY_R16G16B16, and
GDK_MEMORY_R16G16B16A16_PREMULTIPLIED, which
matches the libpng PNG_FORMAT_LINEAR_RGB_ALPHA
format. These formats will be used to provide
linear (gamma-corrected) input in future commits.
GL can upload these formats directly. This is not
only less work for us, it also avoids losing
precision during upload.