Non-linear regression does not support power function

[gkronber]

In the interpreter for SymReg GP we round powers to the closest integer. Models where this kind of rounding is necessary cannot be handled by autodiff. Therefore, we disabled parameter optimization for all models with power symbols.

However, exponentiation with integer powers could be handled easily in autodiff code for parameter optimization by fixing the power. This would be convenient for NLR because polynomials could be input directly instead of rewriting them to via cube(), sqr(), and products.

Backwards compatibility: For SymReg GP it was not possible so far to use the power symbol in combination with parameter optimization (exception). If we change the parameter optimization code to fail when encountering a non-integer power then the code will still fail because this will naturally occur for GP trees.

[gkronber]

r17817: added support for integer powers to parameter optimization

[mkommend]

Reviewed r17817.

[gkronber]

r18015: merged r17817 from trunk to stable