[MLOps 1.5] Expand the built-ins: NLP

[dbogunowicz]

Feature description

• Added built-in functions: 'mean_score, percent_zero_labels (token classification) answer_length, answer_score, answer_found (question answering), sequence_length (general nlp function)`.
• Added exhaustive unit tests.
• Cleaned up the importing structure to avoid circular dependencies

Coverage:

As specified in the PRD:

nlp_sequence_length -->  covered by "sequence_length "
nlp_percent_unk_tokens --> not covered; see comments below
nlp_answer_found_qa - 1 if yes, 0 if no --> covered by "answer_found"
nlp_answer_length_qa - integer length --> covered by "answer_length"
nlp_score_qa - 0-1 float representing confidence in answer --> covered by "answer_score"
nlp_percent_label_0_token_classification - 0-1 float --> covered by percent_zero_labels
nlp_mean_score_token_classification - 0-1 float --> covered by  "mean score"

Few Comments:

1. We still need to think about a good naming framework, to avoid ambiguity and possible dupes of the functions naming across integrations. Let's discuss it.
2. For computing the percentage of the unknown tokens, I am afraid that currently we are heavily constrained. This information is present during the tokenization process but is not available inside the engine_inputs. Hence, we are currently not able to quickly compute this value in a simple, monadic function.

[bfineran]

LGTM for now, but absolutely need to get something in place to handle item-wise functions

[KSGulin]

LGTM!