fnllm Updates

[darthtrevino]

• Add type-checking blocks to most files to improve load time
• Add audit fields 'created', and 'updated' in cache entries to support cache clearing with TTLs
• Only emit usage metrics on cache misses to improve model consumption accuracy.

fixes #325

[andresmor-ms]

Regarding load times... Have you considered looking into the init files and separate by modules? for example, leave the Limiters stuff in fnllm.limiters, and exposing less things into the fnllm package?

[darthtrevino]

That's not a bad idea, but it's probably worth a longer discussion on how to break that apart. We could just expose top-level types from fnlm, or we could expose nothing; let's address breaking down the barrel-files in a future PR

[AlonsoGuevara]

Just to add on top of Andres' comment. The benefit from type check imports isn't that significant, surely improves time (and Ruff checks them) but the bang for the buck is minimizing how much you're exposing on your top level modules. If everything rolls up to top level, you're basically loading everything and not benefiting from how the Python interpreter lazy loads this stuff.

We did this on graphRAG and load time went from 2 minutes to seconds.