Support declarative row and column filtering

[JacobHayes]

• Support declarative row-wise filters (col X = "..." or X in (...)) of input partitions in the .map method (filtering per (input, output) pair, not input alone), which can be driven by per-partition Statistics the user defines for the Artifact
  • These are orthogonal to column-wise selections, which are defined in the .build method
• View loading logic is expanded to apply these row and column filters in the best way it can (eg: loading from BQ SELECT <subset> w/ WHERE, Parquet reads subset of columns w/ ddf filtering)
• Compared to very granular input partitioning, this:
  • has less overhead (fewer upstream partitions to track)
  • has less precise invalidation (less granular upstream partitions)
  • maintains "small" inputs to the build steps

The # of build tasks is still upper-bounded by the # of output partitions or other concurrency limits.