Cost-based optimizer: Implement simple cost model that demonstrates benefits with NDS queries

[andygrove]

Closes #2718 and #2717

Changes in this PR:

• Implements a simple cost model by specifying default CPU and GPU operator costs, and implements special handling for ProjectExec and UnionExec
• Fixes a bug where writes would fall back to the CPU because the optimizer considered that AdaptiveSparkPlanExec was a CPU plan ([BUG] CBO falls back to CPU to write to Parquet in some cases #2718)
• Fixes a bug where joins were sometimes falling back to CPU ([BUG] CBO forces joins back onto CPU in some cases #2717)
• Added trace logging showing the costs calculated by the cost-based optimizer, and the actions it takes. This makes it possible to debug a query and see what the optimizer is doing.

Real-world benefits of this PR with NDS benchmarks:

• 20+ queries no longer fall back to CPU for writes
• q6: Avoids moving to the GPU for a trivial projection followed by a shuffle
• q11 & q74: Avoids moving to the GPU for a trivial projection following a CPU SortMergeJoin. The next operator moves to the GPU anyway, so this isn't of much benefit overall.
• q23a: Avoids moving to the GPU just for a union
• q58: Avoids moving to the GPU for a trivial projection between two CPU SortMergeJoins.

WIP because I can't test everything locally and need time on environments to do full runs at scale after code changes, to confirm no regressions. I also need to figure out how to write good tests for the current capabilites.

[jlowe]

build