Deprecate ICNS (width, height, scale) sizes in favour of load(scale)

[radarhere]

Normally, an image size is (width, height). However, the ICNS format allows a third dimension to be added to this, scale, that only exists before loading.

https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html#icns

sizes
A list of supported sizes found in this icon file; these are a 3-tuple, (width, height, scale), where scale is 2 for a retina icon and 1 for a standard icon. You are permitted to use this 3-tuple format for the size property if you set it before calling load(); after loading, the size will be reset to a 2-tuple containing pixel dimensions (so, e.g. if you ask for (512, 512, 2), the final value of size will be (1024, 1024)).

This is a brief exception to the rule that image sizes have two dimensions. As such, it complicates type hinting, not just in our implementation, but also conceptually for the user - it would be confusing to explain that images might have three dimensions, but only for one format and only briefly.

This PR suggests that we deprecate this third dimension, and instead take an idea from EPS.

https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html#loading

If Ghostscript is available, you can call the load() method with the following parameters to affect how Ghostscript renders the EPS.

scale
Affects the scale of the resultant rasterized image. If the EPS suggests that the image be rendered at 100px x 100px, setting this parameter to 2 will make the Ghostscript render a 200px x 200px image instead.

So instead of

im.size = (16, 16, 2)
im.load()

the user can now

im.size = (16, 16)
im.load(2)