Re-integrate better into mlx_lm

[uogbuji]

Our upstream llm_structured_output, which has since been vendored in, was developed prior to some core mlx/mlx_lm features, and lots of interface changes. It's a bit stale, and we're pretty clearly tripping over wires everywhere.

Time to rewrite the code, with primary locus of https://github.com/OoriData/Toolio/blob/main/pylib/schema_helper.py

[uogbuji]

Note: we'll want to keep an eye on mlx_lm's new speculative decoding, and decide whether to wire in that option.