Promote vision functions out of beta

[jlowin]

Vision is no longer in beta, as it is officially supported on both GPT-4-Turbo and GPT-4o. That means we can move Marvin's vision functions out of beta as well.

Happily, the final versions of the API are much more powerful than the ones we built the original vision functions for. Marvin's original, beta implementation of the various vision-enhanced required two calls to the LLM: one to caption the images with an eye toward capturing detail relevant for the processing task at hand, and a second to take the resulting text and process it (e.g. for classification, casting, etc.).

The final API can do that in a single pass, so I've modified all the classic Marvin functions (cast, classify, extract) to accept image inputs in addition to strings (or, most interestingly, a mix of images and strings!). These work well and are much faster (almost twice as fast!) than the two-pass beta versions.

[jlowin]

Note: this is technically a breaking change since it replaces e.g. marvin.beta.classify with marvin.classify, but that's why those functions were in the beta namespace to begin with.

[zzstoatzz]

✨