Improve handling of randomness in multistart sampling and documentation

[janosg]

This enhancement was suggested by Johannes Schmieder

Current Situation

The default sampling method in multistart optimization is a non-randomized sobol sampling. This is unexpected for users and a problem if someone runs multiple multistart optimizations that have different initial samples.

Describe the solution you'd like

Use a scrambled sobol sequence or some other sampling method (e.g. latin hypercube) as default.

Additional enhancement

Make it clearer in the documentation of multistart optimization how many optimizations will be run and how the settings influence this.