Separate out dataset and evaluator

[rajasbansal]

Currently, the dataset accepts the config object and needs this object to create the ground truth labels and run evaluation. However, the dataset should not need the config object and can accept arguments to do this instead.

Create a separate evaluator class will simplify this and remove dependency of config. It is awkward to initialize the dataset with a labeling config i.e dataaset = AutolabelDataset('test.csv', config)

This is part of a larger goal of making all internal classes config independent.