Improve controlled generation support

[ealmloff]

Currently, Floneum allows you to set a grammar that text generation should conform to. This is useful, but we could take it farther:

Text generation could be represented as a tree of words. The current position in that text is like a cursor in that tree.

In the rust API, we could allow users to either work with the raw cursor API, or use a validation callback:

let session = ModelInstance::new(ModelType::LlamaSevenB);
session.generate_validated(|text| text.len() < 5 && text.chars().all(|c| c == '*');