investigate sixel quantization speedup via final pass

[dankamongmen]

Throughout sixel quantization, I kept harping on the importance of a single pass. During relaxation, we "unzip" and "rezip" various bands when splitting them, to avoid doing a final pass over all the data. In doing so, though, we do several passes (in many cases). Instead, simply perform relaxation, and then do a final pass fitting the sixels. This also leaves us in a good position to replace the quantization algorithm with another one, rather than being wedded to relaxation.

[dankamongmen]

Given that alacritty is likely to have the graphics branch land soon, and the general increased interest in terminal bitmaps, I think it would be wise to do this sooner rather than later. We don't want someone to go benchmark us and publish an unflattering view. This is probably the next thing to work on, actually.

[dankamongmen]

note that libsixel is back in play thanks to @ctrlcctrlv's recent fork. so maybe we needn't bother with this ourselves.

[ctrlcctrlv]

@dankamongmen this is probably going to be something you need to contribute, or at least give me more detail as to how to write it. Right now I'm mostly focused on adding important missing features, not optimizing what's already there. Without transparency it's not useful for kitty at all, for example. Also the inability of the decoder to handle multiple palettes is an issue.

[dankamongmen]

@dankamongmen this is probably going to be something you need to contribute, or at least give me more detail as to how to write it. Right now I'm mostly focused on adding important missing features, not optimizing what's already there. Without transparency it's not useful for kitty at all, for example. Also the inability of the decoder to handle multiple palettes is an issue.

oh sure. what i meant is that libsixel is faster than my quantization right now, at least for most images. if i move to libsixel, it ought be an immediate performance win.

[dankamongmen]

I think we'd be wise to improve sixel performance before releasing 3.0.0. Prioritizing this.