Improve grouping performance via better vectorization in accumulate functions

[alamb]

Is your feature request related to a problem or challenge?

The inner loop of many accumulators, such as Avg, is implemented in NullState::accumulate (source link), which has specializations for nulls and filters.

I think it would be possible to help the compiler make even more efficient code / vectorize the inner loops of the accumulators.

Describe the solution you'd like

Study and optimize the implementation of NullState::accumulate and related functions and make them faster

Specifically, I think being clever about null mask iteration or updating null state more efficiently, we could perhaps make the aggregates faster

Benchmarks to drive some of this work:

benchmarks/bench.sh run tpch_mem
benchmarks/bench.sh run clickbench_1

Describe alternatives you've considered

Here is one approach (that didn't seem to make things better enough): #6954

Additional context

No response

[alamb]

I think this is done now via https://docs.rs/datafusion/latest/datafusion/logical_expr/trait.GroupsAccumulator.html