[FEA] Make row count estimates available to the cost-based optimizer

[andygrove]

Is your feature request related to a problem? Please describe.
In order to implement a real cost model in the cost-based optimizer, it is important to have an estimate of the input row count for an operator. For example, it might be worth moving to GPU for a compute-intensive operation on millions of rows, but not for just a few rows.

Describe the solution you'd like
Spark already contains logic for estimating data sizes (but not row counts) for a logical plan (SizeInBytesOnlyStatsPlanVisitor) and I propose that we based the solution on that logic.

We should also leverage actual row counts when they are available from executed query stages when AQE is on.