Add backend kwargs for pgml.embed

[kczimm]

This PR originally intended to fix #1169. There are some models that do not work with InstructorEmbedding and sentence-transformers but rather should use transformers.AutoModel. Here, we introduce a backend field to the kwargs argument which allows you to opt-in to using the transformers.AutoModel.

Now, you should be able to do the following:

SELECT pgml.embed(
    transformer => 'jinaai/jina-embeddings-v2-base-en',
    text => 'Dynamical Scalar Degree of Freedom in Horava-Lifshitz Gravity"}',
    kwargs => '{"trust_remote_code": true, "backend": "transformers"}'
);

As an aside, I moved the usage and caching of embeddings into Rust and out of Python.