Crop images within 3% of device aspect ratio

[axu2]

Because filling the screen with double page spreads looks amazing on the new Kindle Scribe with its 4:3 aspect ratio. (even though the images are downscaled to 1920x1440 per

• Kindle Scribe mobi limited to 1920 height #460
• fix Add Page Fill Option #391

Also, replaces previous code that could potentially resize images twice, leading to quality loss, instead just using library functions to do it in one pass.

The problem becomes more serious the smaller the edge is.

Here is what the maximum crop looks like:

[axu2]

On further inspection I think this PR is big enough. I recommend squashing my commits.

[darodi]

@axu2

forced pushed a rebase on master with squash and some refactoring

[axu2]

Ah, I meant squash the commits during merge, not in the PR itself. Anyways, just for convenience here are my test images:

[axu2]

Since you already started refactoring, I may as well add the cbz case as well from #471

[axu2]

While we are at it, maybe we should remove the BICUBIC filter. Looks like in the past BICUBIC was better for upscaling but not anymore. This might be a separate PR/logical step though. I personally don't use upscale on the Kindle Scribe since the Scribe will fill the screen no matter what the original size was.

Filter	Downscaling quality	Upscaling quality	Performance
Resampling.NEAREST			⭐⭐⭐⭐⭐
Resampling.BOX	⭐		⭐⭐⭐⭐
Resampling.BILINEAR	⭐	⭐	⭐⭐⭐
Resampling.HAMMING	⭐⭐		⭐⭐⭐
Resampling.BICUBIC	⭐⭐⭐	⭐⭐⭐	⭐⭐
Resampling.LANCZOS	⭐⭐⭐⭐	⭐⭐⭐⭐	⭐

https://pillow.readthedocs.io/en/stable/releasenotes/2.7.0.html
https://pillow.readthedocs.io/en/stable/handbook/concepts.html#filters-comparison-table

[darodi]

Lanczos might produce slightly more ringing than bicubic.
We should try and check if the difference in quality is negligible and what is the impact on performance.
For a large image collection, if the processing time is multiplied, it changes a lot (see mozJpeg for example)

[axu2]

Yea definitely a separate PR then, if we decide to change. I personally think it's probably better to not change.