Enable specifying custom columns and values for splitting.

[justinxzhao]

Currently, Ludwig's dataset splitting requires either:

• Splitting is disabled completely
• Dataset is split randomly into train, validation, and test, according to preprocessing.split_probabilities
• Dataset is split according to a special metadata column in the data, split, with a fixed set of special values to do the association: 0: train, 1: validation, 2: test.

We should extend this API to enable users to customize splitting by other columns and values.

preprocessing:
    split:
        (moved) force_split: false
        (moved) split_probabilities: [0.7, 0.1, 0.2]
        (new) split_column: split # Name of column that should be used for splitting
        (new) train_values: [0] # Values in the split_column that should be associated with the training split
        (new) validation_values: [1] # Values in the split_column that should be associated with the validation split
        (new) test_values: [2] # Values in the split_column that should be associated with the test split

Note: we may want to revisit this API if/when we support multiple test sets.