FEAT: Implement two level dispatcher for execute_node

[icexelloss]

What is the change

This PR implements a TwoLevelDispatcher and replace the multipledispatch.Dispatcher with TwoLevelDispatcher to define ibis.pandas.execute_node

The original dispatcher is found to be slow when
(1) dispatch the first execute_node
(2) dispatch after new signature for execute_node is registered (this happens when user uses UDF)

In this alternative dispatcher, neither of these two issue appears. For details, see docstring of TwoLevelDispatcher

How is the change tested

• ibis/pandas/tests/test_dispatcher.py is added
• Also replies on existing pandas backend tests.
• Manually tested tracing and verified tracing still works.

Benchmark

A benchmark is performed by running all the pandas backend tests and record "time to dispatch" for all signatures.

Here is the top 10 slowest dispatch in the original implementation (with the corresponding time in the new implementation):

Here is the top 10 slowest dispatch in the new implementation (with the corresponding time in the original implementation):

[jreback]

conceptually looks fine. the perf of ops seems slighy slower when they are not the worst case ops (cast float64, series). is this noticeable? can we specifically register UDF ops with the 2-level and leave everything else with the 1-level (does this make it way more complicated)?

• i think this makes sense to push upstream to multidispatch as well (ok in here until / unless this patch is accepted & released)

[icexelloss]

I am leaning towards not optimize this further:

• The performance is good enough. Even for an execution that triggers all of the first 10 slowest dispatches, the overhead is still < 0.2s. It's not perfect but I don't think it's noticeable either.

• Support both 1-level and 2-level in a single dispatcher does make it more complicated. - for a given signature, we would first try to decide whether it's registered in 1-level or 2-level dispatchers, which by itself can be quite expensive.

• I don't want to commit to further optimization before researching if other "multi dispatch" libraries are better suited for ibis. There is also multimethod and multimethods (yes, they are different) library that does similar things. This can potentially be more work so I'd like to keep it out of this PR.

[jreback]

may want to consider a method for this on the dispatcher itself? (other wise this leaks into the impl).

[icexelloss]

Added delitem for this

[icexelloss]

Thanks! @jreback