add a round-up option to shrink

[jcupitt]

resize does shrink for the int part, then reduce for any fractional residual.

The shrink stage rounds image dimensions to the nearest pixel, so it can trim up to 0.5 px from the right and bottom edge. This means that the second reduce phase can have some missing information compared to operating directly on the unshrunk image.

As an improvement, resize could request round-up from shrink, and then pick a scale factor for reduce that still gave the correct sized output, but no longer lost the 0.5px of information in the intermediate.

See libvips/pyvips#148 for the original context.