refactor: extensible function registry

[universalmind303]

I didn't want to create 1 massive undecipherable PR, so I wanted to submit this first.

This ties a FunctionRegistry instance to a session instead of the previously global FUNCTION_REGISTRY.

Right now, the entries are written to both the catalog and the function registry.

I'd prefer writing to just one or the other, but havent quite figured that out yet as our udfs are not serializable so they cant be written directly to the catalog.

Some other the notable changes so far

• some catalog changes to allow for adding new user defined functions to the catalog.
• internally to expose a new method register_function on the session.
• rework cosine_similarity to the new datafusion udf impl pattern for easier usage with the new register_function method.

I manually tested this by removing cosine_similarity from the builtin functions and registering it manually

let mut sess = engine
    .new_local_session_context(SessionVars::default(), SessionStorageConfig::default())
    .await?;

let udf = Arc::new(CosineSimilarity::new());
sess.register_function(udf).await?;

This works as expected in local sessions only. We'll have to modify the planner to handle udfs in a hybrid context at a later date.

---

Next steps.

I'm going to try to integrate the work I did with this datafusion <--> polars POC in a subsequent PR.

https://github.com/universalmind303/polar-fusion

[tychoish]

seems reasonable.

• this needs automated tests.
• should our other functions migrate over as well?

[universalmind303]

this needs automated tests.

There's really nothing to test so far that's not already covered by all of tests. This is all internal. If things were to have broke, our existing tests would have caught known any regressions.

should our other functions migrate over as well?
yeah we have an issue open for that already.

[scsmithr]

Could we add an SLT for the cosine similarity function just to make sure it's all wired up correctly (even if it needs to be skipped for rpc/flight).

[universalmind303]

Could we add an SLT for the cosine similarity function just to make sure it's all wired up correctly (even if it needs to be skipped for rpc/flight).

@scsmithr we already have a bunch of tests for it. testdata/sqllogictests/functions/cosine_similarity.slt

[tychoish]

Let's hold off on merging this until we decide if #2715 is worth cutting a quick 0.9.1 over.

[greyscaled]

ps: this is cool!