Penalty rewards

[p-ferreira]

This PR seeks to propose a penalty mechanism with:

- Penalty rewards: A new family of function models that will act as penalty functions accordingly to the established definition of the function. This PR also introduces a task-criteria schema that enable criteria assignment to a given task through prompt definition and code validation.

Tasks done so far:

• add generic task and criteria schema that can be easily expanded
• add SummaryTask, QuestionGenerationTask and QuestionAnswerTask
• adds MatchLengthCriteria with support to characters, words, sentences and paragraphs
• adapt current workflow to include task validation
• adapt current workflow to not concatenate previous answer and questions
• add penalty functions
• incorporate legacy task validator reward model functionality into new KeywordMatch penalty function
• add SentenceMatch penalty function

TODO (Once current design is approved):

• reorganize and properly document the code generated
• add unit tests for implemented functionality

Wandb run with penalty functions implemented:

• https://wandb.ai/opentensor-dev/openvalidators/runs/7rahrixe
• https://wandb.ai/opentensor-dev/openvalidators/runs/s7fl9hea

[Unkownman086]

With the new prompting style you receive much worse ratings from the RLHF model.
I have tried with same context and i am able to measure from average GPT-3.5, vicuna response much worse rating on RLHF model, when normalized ≈ 30% worse.

I whould think the RLHF model doesn't do well, when the question is at the start of the prompt

[steffencruz]

Please make penalty_scale_factor in prompting/validators/criteria.py a gaussian as i showed in the plot.