Minor reduction improvements

[andrews05]

This started off as minor a bug fix but then expanded into a larger refactor.

The bug occurred when a GrayscaleAlpha image with a bKGD chunk tried to reduce to Indexed. The bKGD conversion was assuming RGB and would fail the reduction entirely if the chunk wasn't 6 bytes. It now correctly reduces a Grayscale bKGD and thereby allows the reduction to Indexed. (My fault, since I added the GrayAlpha to Indexed support)

However, due to the new found ability to reduce such images to indexed, "issue-60.png" ended up regressing slightly as the indexed version was actually larger than the gray alpha. Ultimately I ended up completely refactoring the reduction and evaluation sequence, combining the perform_reductions function with reduce_color_type (I've put the code in the reduction mod, mainly to reduce the amount of code in lib).
The result is a bit tidier and has a few notable outcomes:

• Grayscale can now be reduced from 16 bit to depths lower than 8, where before they would stop at 8. I've added an extra test for this.
• Conversions to Indexed can now be evaluated at 8-bit as well as a lower depth where applicable. This can help with small data where 8-bit is sometimes better (e.g. some images in Leanify tool provides better lossless compression than Oxipng. #418). Fixes Implement possible size optimizations from Bit Depth from TruePNG #73.
• Certain reductions, such as 16-to-8 and rgb-to-gray, are performed unconditionally with no evaluation, as these are just byte removals. These reductions then become the baseline that subsequent reductions are evaluated against, which again can help with small data (previously in a chain of RGBA->RGB->Indexed, the RGB in the middle would not be considered). Checks off item 4 in Implement possible size optimizations from Color Type from TruePNG #72.

The end result of all this is... "issue-52-04.png" now compresses 3 bytes smaller 😂
Performance is essentially unchanged (I did measure a fractional improvement at level 0).

[andrews05]

Well, this turned into a larger refactor than I had planned 😆

@kornelski I ended up removing the extra 4-bit evaluation. I managed to find some images which did prefer 4-bit to 2-bit, but (perhaps due to improvements since that code was added) they were actually best at 8-bit (or even rgb). So I think it's not providing any benefit anymore, but do let me know if you disagree.

[kornelski]

can reduce_to_palette be moved before line 80 to try reduced_palette once?

[andrews05]

The two calls to reduced_palette are mutually exclusive: The earlier one will run if the image is already indexed (the function will simply return if it isn't), while this one will run if we've just converted to indexed.

[shssoichiro]

Thanks, looks good, just needs a rebase 👍

[andrews05]

@shssoichiro Thanks, done 🙂