Optimize getting the palette index for palette compression using SIMD.

[IntegratedQuantum]

The compiler is poor at optimizing search loops like this, since it doesn't have enough information (it must not access the memory beyond the last entry and thus is stuck iterating element by element).

So I decided to try and throw some SIMD on this, and the easiest way to use SIMD is possible is to align the vector and blindly access memory beyond the length limit, and it actually is not a lot more complex than the non-SIMD code (see godbolt).

I also did some measurements, but getting the palette is not a common operation, and thus overall performance impact is below 1%, the given function measured alone though is about 17-50% faster depending on the use case.
But to be honest this is mostly about fixing the worst-case performance (which of course I did not care to measure :​P)

• cleanup, move function, orelse, maybe some name improvements

fixes #318

[Argmaster]

How did you measure performance impact?

[IntegratedQuantum]

I looked at it with a sampling profiler, not terribly precise I know. Maybe I'll make a better worst-case benchmark for this in the future.