Insert buffer converters for TypedImperativeAggregate

[sperlingxx]

Current PR addresses the last task of #2916, supporting the aggregation buffer conversion between CPU and GPU format for TypedImperativeAggregate functions. With the ability to insert buffer converters, we can handle TypedImperativeAggregate functions running across CPU and CPU in runtime. It indicates that we don't need to fallback the entire Aggregate stack to CPU once one stage need to fallback when the Aggregate contains TypedImperativeAggregate functions.

The general idea is to create buffer converters and bind them to certain physical plans during preColumnarTransitions. And integrate these buffer converters into RowToColumnar/ColumnarToRow transitions as pre-processing/post-processing projections during postColumnarTransitions. And this idea works even when AQE is on. To adapt AQE, we leverage TreeNodeTag to cache temporary information, including: buffer converters, and some meta data.

For better understanding, let's walk through the entire procedure of inserting buffer converters:

1. binding buffer converters to certain plans (preColumnarTransitions)
   1.1 collecting all stages of Aggregation which contains TypedImperativeAggregate functions
   The binding procedure is triggered in GpuTypedImperativeSupportedAggregateExecMeta.tagPlanForGpu if wrapped plan
   is the final stage. At first, we collect all stages as what we do for associated fallback.
   1.2 filtering stages who need buffer converters to fill the data gap with their child
   1.3 creating buffer converters with filtered stages
   We add two new interfaces on GpuTypedImperativeSupportedAggregateExecMeta: createCpuToGpuBufferConverter and createGpuToCpuBufferConverter.
   1.4 binding buffer converters into certain plans
   The plans who carry the buffer converters are the CPU plans (can not be replaced) located right before/after the potential R2C/C2R transitions.
2. materializing RowToColumnar/ColumnarToRow transitions with buffer converters (postColumnarTransitions)
   We add extra field preTransitions/postTransitions to GpuRowToColumnar/GpuColumnarToRow, in order to insert projections like buffer converters for TypedImperativeAggregate.

[sperlingxx]

build

[revans2]

I am a little concerned that this code appears to assume AQE is on all the time. It also looks like we cannotl have a TypeImerativeAggregation that does not use transitions. For percentile approximate it is going to use a very different algorithm and only supporting transitions is going to be very hard. I would prefer to keep the old code if we have to pick just one way to do it.

[sperlingxx]

build

[sperlingxx]

I am a little concerned that this code appears to assume AQE is on all the time. It also looks like we cannotl have a TypeImerativeAggregation that does not use transitions. For percentile approximate it is going to use a very different algorithm and only supporting transitions is going to be very hard. I would prefer to keep the old code if we have to pick just one way to do it.

I brought back the old codes for "associated fallback". After that, we will judge whether all TypedImperativeAggregate buffers across CPU and GPU are available to be converted in runtime. If so, we insert buffer converters. Otherwise, we just fall back the entire Aggregation stack.

[revans2]

It looks good. I just have a few nits, but I am happy to let it in without any changes

[sperlingxx]

build

[tgravescs]

it looks like this is missing the databricks shim updates for GpuColumnarToRowTransitionExec. If you are touching all the shims please think about the databricks ones as well.

[tgravescs]

I'll put up pr to fix

[tgravescs]

Sorry it looks like this may have just been in my branch with various build changes, apologize