Flatten logging parameters to improve readability

[nyanp]

In run_experiment, model_params and fit_params are stored as string to distinguish them from other parameters and it makes difficult to compare params across experiments.

It seems better to flatten the dictionary as in the PR below.
mlflow/mlflow#1863

Before

{
    "fit_params": "{ early_stopping_rounds: 100 }",
    "model_params": "{ max_depth: 3, objective: \"binary\" }",
    "algorithm_type": "lgbm"
}

After

{
    "fit_params.early_stopping_rounds": 100,
    "model_params.max_depth": 3,
    "model_params.objective": "binary",
    "algorithm_type": "lgbm"
}