Avoid side-effects in constructors of data analysis solutions

[abeham]

In #2238 it was decided to strip the grammar off the symbolic expression tree when creating a new solution due to memory optimization reasons.

The changes however lead to a side-effect in the constructor of e.g. SymbolicRegressionSolution. A modification of the tree is not expected or apparent. It would be more friendly to the user of the API if such a kind of optimization is offered via a separate method than the default way through the constructor. Other possibilities would be to use a clone of the model and of the tree in the solution.

Consider the evaluation function of the following programmable problem

    public double Evaluate(Individual individual, IRandom random) {
      var model = individual.SymbolicExpressionTree();
      var model = new SymbolicRegressionModel("Target", model, new SymbolicDataAnalysisExpressionTreeInterpreter());
      var solution = new SymbolicRegressionSolution(model, problemData);
      return solution.TrainingMeanAbsoluteError;
    }

Optimizing this e.g. in a GA results in a NullReferenceException in the SubtreeSwapingCrossover.