Calculate `p_mutate` using multi-armed bandit

[Liam-DeVoe]

Originally: #177 (comment)

I think we want to track (whatever estimator state we use for estimating behaviors/fingerprints per second, currently "inputs since new coverage") independently for blackbox and mutator, and do a multi-armed bandit optimization between the two to generate p_mutate. This will cleanly inherit any thompson sampling improvements.