See if we can avoid returning scan batches that contain no data

[nicklan]

In debugging #233 , I realized we have a case where the first batch of data does include and Add file, but that it's filtered out due to not matching the predicate that's been pushed down.

We can't just check if all add files are masked, because we just have an EngineData, but perhaps the visitor could remember the indexes of all the add files, and if they are all filtered out, just go on to the next batch. We already do this in log_replay if the entire selection vector is false, but in this case, it wasn't because the other data item was a Remove.