Recompute learning rates depending on mu

[nikohansen]

Ideally, the code should "adapt" to varying $\mu$. For example, when we only find a few feasible solutions, or when we inject a feasible solution to guide the search, we may want to be able to make a "premature" update that still works without, say, losing eigenvalues in the covariance matrix.

This means that tell should recompute the internal parameters when necessary, possibly based on an optional input argument $\mu$.

Next steps:

• review and specify the input to be used ($\mu$, len(solutions)) and the settings to be done
• review and list the code/places which need to be changed

See also issue #7.