Optimize color counting

[kornelski]

I've seen a panic in most_popular_edge_color, so I've made it more resilient.

[AlexTMjugador]

Thank you, these changes look sensible. Do you have any clue on why len can be outside of the expected range in the affected code?

[kornelski]

I only have a core dump without the image, so I don't know exactly.

A common tricky case that may be related is pixels having higher values than palette length.

It also could have been 1x1 interlaced image.

[andrews05]

Thanks for this @kornelski. I'll take a closer look later and try to figure out what would actually cause this, but I'm surprised it can happen since:

1. The image should have already been cleaned from any palette issues at this point.
2. Interlaced images are explicitly excluded from the co_occurrence_matrix sort methods.

[AlexTMjugador]

Both of your comments sound good! I think this PR could be merged as-is, as it's not exclusive with actually figuring out what's causing the bad len, but I'll leave the final choice to @andrews05, given Andrews is more familiar with this code.

[andrews05]

I managed to reproduce a crash at that point. @kornelski was right - it can happen when palette entries are missing.
The reduced_palette is supposed to fix any such issues, but it won't if the palette was one entry too short. I'll make up a separate PR with a test case to fix this.