API, Core: Update inclusive metrics evaluator for extract and transforms

[rdblue]

This updates InclusiveMetricsEvaluator that uses column stats to skip data files during scan planning.

The evaluator was implementing the older BoundExpressionVisitor interface that only supported BoundReference and not other BoundTerm instances like BoundTransform. After #12304, BoundExtract also needs to be supported. This PR includes #12304 and will be rebased when it is merged.

Filtering works for transformed values when the transform is order preserving. If it is not order preserving (like bucket) the bounds cannot be used.

Most of the changes to BoundExpressionVisitor are to produce the lower and upper bounds values that are tested:

• For BoundReference, deserialize the bound from the correct lowerBounds or upperBounds map (moved to parseLowerBound and parseLowerBound
• For BoundTransform, deserialize the bound and transform it if the transform is order-preserving (in transformLowerBound and transformUpperBound)
• For Extract, deserialize the bound as a map from field name to VariantValue, then convert the value to the internal representation (in extractLowerBound and extractUpperBound)

This adds new test suites for the new BoundTerm cases that are supported:

• TestInclusiveMetricsEvaluatorWithExtract tests variant cases
• TestInclusiveMetricsEvaluatorWithTransforms tests transform cases

[rdblue]

If the bounds are in metadata, they are trusted. Bounds should be missing in cases where there are variant values with incompatible types.

[rdblue]

This serialization is temporary and will be replaced with a well-defined variant serialization.

[rdblue]

This needed changes to TestSparkScan because adding support for transforms filters out additional files in unpartitioned tables.

[rdblue]

This now relies on #12374 to move the serialized variant classes into API so that InclusiveMetricsEvaluator can use them for deserializing bounds.

[rdblue]

Thanks for reviewing, @danielcweeks!