[ML] Refactor scaling of input values in CBayesianOptimization

At the moment we are normalizing the input values by 1/(max-min) whenn adding them to the GP. We do it to improve the numerical stability of the kernel matrix. However, it leads to an logically inconsistent state, where the stored values can be outside of the min/max boundaries.

We should investigate if we can move the normalization coefficient into the theta1 kernelParameters to avoid potential errors.