Avoid using AST on inner joins and avoid coalesce after nested loop join filter

[jlowe]

Fixes #3736.

@abellina and I looked into the nested loop join performance degradation and there were primarily two issues:

1. Computing the AST during the join was particularly expensive, resulting in approximately 4X performance improvement for the microbenchmark nested loop join if we revert back to generating a full cross join output and then post-filtering the results.
2. The GpuFilterExec extracted from the join in Add filter in query plan for conditional nested loop and cartesian joins #3105 was causing a subsequent batch coalesce to occur that wasn't there before. Traces showed contiguous_split was quite expensive on the post-join output for this query. Updating the post-join filter to not coalesce batches resulted in another approximate 4X speedup from the previous change.

This updates nested loop joins to avoid using the AST if a post-filter can be used and also updates the post-filter injection code for joins to specify no coalesce after goal on the filter node.

[jlowe]

build

[jlowe]

build

[jlowe]

Marking this as a draft until NDS performance runs are complete to verify there are no regressions.

[jlowe]

NDS saw a perf regression in q72, the coalesce after the post-join filter turned out to be crucial in that case. Refined this to only avoid the coalesce on a nested loop join due to the microbenchmark results, filed #3749 and #3750 to track fixing this in the general case.

[jlowe]

build

[jlowe]

NDS and cross join microbenchmark results on latest patch look good, marking this ready to merge.