feat: Deduplicating recursive CTE implementation

[Tpt]

Rely on aggregate GroupValues abstraction to build a hash table of the emitted rows that is used to deduplicate

We might make things a bit more efficient by rewriting a hash table wrapper just for deduplication, but this implementation should give a fair baseline

Which issue does this PR close?

• Closes Support deduplicating UNION in recursive CTE #18140.

Rationale for this change

Implements deduplicating recursive CTE (i.e. UNION inside of WITH RECURSIVE) using a hash table. I reuse the one from aggregates to avoid rebuilding a full wrapper and specialization for types. Each time a batch is returned by the static or the recursive terms of the CTE, the hash table is used to remove already seen rows before emitting the rows and keeping them in memory for the next recursion step.

What changes are included in this PR?

Reusing GroupValues trait implementations inside of RecursiveQueryExec to get deduplication working.

Are these changes tested?

Yes, some sqllogictests have been added, including ones that would lead to infinite recursion is deduplication where disabled.

Are there any user-facing changes?

No

[tobixdev]

From my perspective this is a very nice and concise solution to the problem.

Furthermore, from my understanding this should also correctly terminate the recursion as only each unique row is pushed into the WorkTable and at some point (as it can be seen in the closure example) this will reach a fix point.

What I am also thinking about is test coverage. My gut feeling says there should be some test cases in the SQLite test suite that cover distinct recursion. Would this cause the extended test suite to fail? Ideally, this solution passes all these test cases now! 🥳 However, I am a bit unsure how this is setup currently.

Thank you!

CAVEAT: I am by no means a DataFusion (nor recurisve query) expert so take my comments with a grain of salt.