WIP: Support epsilon_relaxation and goal relaxation when keep_soft_constraints is True

• Merge !186 (merged) first
• With this implementation, a relaxation of e.g. 1E-6 combined with 100.000 goals/epsilons will also allow one goal to take an epsilon of 0.1 (1E5 / 1E6), with all others remaining zero. That's probably not what we want. Should we add constraints on all epsilons instead (<= ...), or in addition? How would this then be different from the old goal programming (keep_eps_variable = False)?
• Support goal relaxation