[7.x][ML] Hyperparameter importance

[valeriy42]

This PR implements hyperparameter importance calculation using ANOVA decomposition of the Gaussian process model used in the Bayesian optimization routine. The results is recorded in the model metadata structure. For convenience, I record the anova variance as an absolute value as well as a fraction of the total variation. This "relative importance" simplifies the comparison between different parameters.

Additionally, I record the best value for the individual hyperparameter s.t. it can be used in a follow-up job configuration and doesn't have to be looked up separately.

Example output:

"hyperparameter_importance": [
        {
            "name": "downsample_factor",
            "value": 0.35355339059327386,
            "absolute_importance": 0.123,
            "relative_importance": 0.9123
        },
        {
            "name": "alpha",
            "value": 0.010766199582740555,
            "absolute_importance": 0.0123,
            "relative_importance": 0.0321
        }
    ]

In order to verify the analytical formulas, I implemented Sobol sequence random number generation, which leads to lower variance for Monte-Carlo estimations.

Backport of #1627