Add a multiobjective adaptation for DE

[jrapin]

Types of changes

• Docs change / refactoring / dependency upgrade
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)

Motivation and Context / Related issue

Our DE implementation performs really poorly on dynamic functions, which arise when doing multiobjective optimization. This update make it work in this case by using the pareto front information.

What that changes. In case of multiobjective:

• best is sampled in the pareto front if the initial point is not on the pareto already (otherwise, no best)
• a and b are sampled on the pareto front as well if there are more than 3 points there
• the child replaces its parent with probability "the ratio of losses which are lower for the child than for the parent".

How Has This Been Tested (if it applies)

Checklist

• The documentation is up-to-date with the changes I made.
• I have read the CONTRIBUTING document and completed the CLA (see CLA).
• All tests passed, and additional code has been covered with new tests.

[jrapin]

This is indeed much faster to converge: blue with adpatation, orange without, best solution at (80, 100)

[jrapin]

At first sight, it looks roughly equivalent (and maybe more stable) than with mono objective:

• red and orange as multiobjective
• blue and green as monoobjective: one of them is completely lost

[jrapin]

It seems (after few repetitions) to work alright on the new_multiobjective_example benchmarks (note that cigar functions are removed to avoid bad initializations, DE-noadapt and TwoPointsDE-noadapt are without adaptation):

[jrapin]

Consolidated results show a very clear improvement over non-adapted DE, and all other optimizers for that matter (none of which is taking into account the pareto front for the time being).