[FEA] Filter that also drops columns

[jlowe]

A somewhat common occurrence in queries is to see a filter on one or more columns followed by a project that drops the columns that were filtered on (i.e.: the columns were only needed to perform the filter). Currently we're manifesting the result of the filter on such columns only to drop them in the subsequent project. It would save time and GPU memory if we were able to transform the Filter --> Project into a FilterWithDrop that did not manifest the filter result for columns that are no longer needed after the filter.

[revans2]

This is very similar to #8831 we might also want to look into window dropping columns. I know that it can happen, especially if the column is just used to create the window and does not need to continue afterwards. I don't think window would be a performance improvement. We are not going to gather the data for window. But it would be a potential memory reduction by dropping the column slightly earlier. But maybe not that big of a win.

The only other operator I could think of that might drop a column after it is used and we could avoid a gather is hash aggregate, but that would require changes to CUDF, and I don't think it is that common of an operation.