[FEA] Support common subexpression elimination for expand operator

[winningsix]

Is your feature request related to a problem? Please describe.
Expand operator may have multiple Seq[Expression] as below

spark-rapids/sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuExpandExec.scala

Line 91 in 5d08aec

val boundProjections = projections.map { pl =>

where it may exist some common subexpressions across those expression sequence. In Spark, most of such cases can be optimized via column pruning logical rule. But it may fail in some cases to extra expressions as a child project node. As a result, some duplicated computations will happen given it's a forest other than a single tree of expression seq.

Given following thing as an example:

Expand
Projections:
  [
    [rapids_fields, null, null, null, null, 0,
      if (complex_expr) 1 else 0],
    [rapids_fields, null, 0,
      if (complex_expr) 1 else 0],
      null, null, null, null]
   ]

If it failed to be optimized by logical plan, complex_expr will be duplicately evaluated.

Describe the solution you'd like
Introduce a new approach allowing extract project where it can have common sub-expression elimination via tiered project evaluation approach.

With optimization within expand node, we can do something at physical level to fix up cases failed to have expand expressions extracted.

Expand 
  projections:
    [
    [rapids_fields, null, null, null, null, 0, ref#1],
    [rapids_fields, null, 0, ref#2, null, null, null, null]
    ],

Project 
    [ 
     rapids_fields, if (complex_expr) 1 else 0 as ref#1, if (complex_expr) null else 0 as ref#2
    ]

Describe alternatives you've considered
Have a fixup case-by-case when column pruning failed to happen in expand node.