Simplified a lot of code which was unnecessarily generic.
Now using monads to manage the state of the random number generator
rather than passing it around by hand.
Also made some performance improvements, then increased the population
size so more combinations are tried in a similar length of time.
The previous code to calculate the accentDifference for the fitness
considered cases where a=b, so the minimal difference was always 0.
The new code explicitly excludes the index, so a != b (except both
entries were sampled to the same color, but that is what we want to
punish here).
This allows us to reach an even better colour scheme rather than
stopping after a fixed number of generations.
For images which create a good colour scheme quickly, we avoid doing
unnecessary passes which don't lead to much improvement.
