Remove references to sacred

This refactoring removes sacred and implement explicit configuration management using teklia-toolbox instead of relying on Sacred's magic configuration style (variable discovery).

The experiment now supports a list of JSON configuration files as input, merge them, and process them into a single configuration dictionary. There is no more params nor global_params section, even though they are still supported and will be promoted to the root level of configuration.

Here is a sample of the new supported configuration:

{
    "batch_size": 4,
    "bin_size": 20,
    "classes_colors": [
        [
            0,
            0,
            0
        ],
        [
            0,
            0,
            255
        ]
    ],
    "classes_names": [
        "background",
        "text_line"
    ],
    "data_paths": {
        "test": {
            "image": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/test/images"
            ],
            "json": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/test/labels_json"
            ]
        },
        "train": {
            "image": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/train/images"
            ],
            "json": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/train/labels_json"
            ],
            "mask": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/train/labels"
            ]
        },
        "val": {
            "image": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/val/images"
            ],
            "json": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/val/labels_json"
            ],
            "mask": [
                "/home/bastien/dev/teklia/ufcn-datasets/data/DLA/IAM/val/labels"
            ]
        }
    },
    "experiment_name": "iam",
    "img_size": 768,
    "learning_rate": 0.005,
    "log_path": "runs/iam",
    "mean": "mean",
    "min_cc": 50,
    "model_path": "model.pth",
    "no_of_epochs": 200,
    "no_of_params": null,
    "prediction_path": "prediction",
    "save_image": [
        "val",
        "test"
    ],
    "std": "std",
    "steps": [
        "normalization_params",
        "train",
        "prediction",
        "evaluation"
    ],
    "tb_path": "events",
    "training": {
        "loss": "initial",
        "restore_model": null
    },
    "use_amp": false
}