[Hyperopt] Allow default shared parameter search spaces

[arnavgarg1]

It would be helpful to enable using a new defaults keyword within hyperopt to set default parameters for feature groups. This will help add search spaces more concisely for datasets with a large number of features while also reducing the search space to allow for a deeper search during hyperopt.

For e.g., a user should be able to add

"defaults.input_features.text.cell_type": {"space": "choice", "categories": ["lstm", "gru"]}

in their hyperopt search space so that all text input features use the sampled cell_type for that particular trial.

The API for this change will be as follows: defaults.<feature_group>.<feature_type>.<parameter>, where:

• feature_group: either input_features or output_features to distinguish which features to apply this to
• feature_type: the feature type in question, for e.g., text/category/sequence etc.
• parameter: the parameter that one wants to search over for that particular feature type, for e.g., vocab_size for text

Example Configuration:

hyperopt:
    goal: minimize
    parameters:
        trainer.learning_rate:
            lower: 0.0001
            upper: 0.01
            space: loguniform
        defaults.input_features.text.cell_type:
            space: choice
            categories: ["lstm", "rnn", "gru"]
        defaults.output_features.category.vocab_size:
            space: randint
            lower: 4
            upper: 8
        ....
    ....