Feature/embeddings

[JackHopkins]

Adding support to return embeddings for retrieval use-cases in the following syntax:

@monkey.patch
def score_sentiment(input: str) -> Embedding[np.array[float]]
    """
    Scores the input between 0-10
    """

With alignment statements defining contrastive fine-tuning cases in the same syntax of existing align statements:

@monkey.align
def align_score_sentiment():
    assert score_sentiment("This food is disgusting") != score_sentiment("This food tastes great!")

We are storing positive and negative contrastive examples, using the new Bloom filter implementation.

Because we can't use the patched embeddings to improve the quality of the model (AFAIK), we don't save them.

The new terminology for standard patches/align statements is Symbolic, and for embedding patches/align statements it is Embeddable.

There is now an EmbeddingModelManager to handle the creation of prompts and to generate the embeddings.
LanguageModeler is now called LanguageModelManager and it is responsible for repairing and managing the operations of the backend LLM.

High-level LLM providers can inherit from Embedding_API and/or LLM_API to expose embedding and symbolic sampling support respectively.

The next step is to move the low-level OpenAI finetuning logic into something like an SymbolicFinetuningAPI which handles training distilled models.

Conversely, we could also create a EmbeddingFinetuningAPI to handle contrastive finetuning using the examples we are now collecting from Embedding functions.

Currently only OpenAI is supported - and it inherits from both Embedding_API and LLM_API.