Output normalization

[RunDevelopment]

This adds the output normalization I recently mentioned on discord. The main change is that ImageInput no longer normalizes, ImageOutput does the normalization instead. This is more efficient because:

1. A single output can be connected to multiple inputs. Previously, each input would normalize the image, resulting in duplicate work.
2. We can tell the output that the returned image is already normalized (assume_normalized=True).

Only when an image output is not used is the new system less efficient. However, this case is relatively rare and can be made equally efficient with assume_normalized=True for already normalized images. Luckily, this trick can be used for Separate RGBA and Split Transparency.