Extend early stopping callback

[rpreen]

The current EarlyStoppingCallback terminates fitting based solely on the error. However, in many situations a compact rule-set is desired. Allowing XCSF to continue to iterate after the system error falls below $\epsilon_0$ provides more time for the evolutionary algorithm to discover more general solutions and for compaction methods such as subsumption to reduce the number of microclassifiers. Additionally, in some problems and with some rule representations, quite a large maximum population size is necessary to initially discover sufficiently low error solutions before the generalisation process can begin.

The current EarlyStoppingCallback could be extended (or a new one introduced) to terminate after the system achieves $\epsilon_0$ and the rate of population compaction has fallen below some delta.