Added better functionality for label selection

[Vaibhav2001]

Pull Review Summary

Continued from #561. This closes #712

Description

• Now we can select the top K labels to pass to the LLM with label_selection, and specify the count though label_selection_count, and a threshold via label_selection_threshold.
• Enabled caching for storing embeddings.
• Few shot examples will only contain the correct examples which have the provided label. (only works with "fixed" algorithm for now)

Benefits:

• Leads to a similar accuracy as base classification, however leads to a better completion rate when num_labels is huge, as the model prediction does not map exactly with any label (mostly because of large context length).
• Cost reduction as the number of tokens passed to the final model can be small.

Usage:

config = {
    "task_name": "BankingComplaintsClassification",
    "task_type": "classification",
        ...
        "labels": label_descriptions,
        "label_selection": True,
        "label_selection_count": 10,
        "label_selection_threshold": 0.5,
        ...
    },
}

• Can use label descriptions to embed the labels.
• "label_selection_count": -1 makes maximum possible labels equal to len(labels). Defaults to min(10, num_lables)
• label_selection_threshold enables us to only choose the labels where label_score / topScore > threshold for effective filtering. This value defaults to 0.95 (based on experimentation).

Type of change

• Improving existing feature
• This change requires a documentation update

Tests

Ran benchmarking tests for different datasets.

Future Work

• Do not classify if there is only one selected_label. Needs to change the annotation processing in dataset.process_labels.
• Costing for embedding model. However this is typically upper bounded by the base model and is cheaper by a factor of up to 1000.

[rajasbansal]

lgtm - after small fixes