[Enhancement] Custom selector function

So I was on #982 , where weixuanfu mentioned that as for TPOT right now, "mutation could happen on transformer step only in this small-size population and the population became homologous in one or two classifiers with highest scores", and he proposed helping the situation by finding maybe a better mutation method.

I was thinking if we could instead not eliminate the worse performing individuals so harshly, so like balance the selection probability more. This then led to me thinking if it is possible that when setting a config we allow the user to set their own selector function, just as how the scoring function can be customized.

So then lets say we want to be less harsh, instead of roulette wheel with the fitness as probability, we can log the fitness values to make probability of more and less fit individuals closers, or any other way really.

Alternatively if anyone knows how to start approaching this enhancement, please enlighten me so I may try to do it myself and send a pull request.

Thank you.