Clean up physical layer

[aannleax]

The implementation of column scans is a bit inelegant. The trait ColumnScan defines the functions pos and narrow which only make sense for those column scans which directly iterate through a column. However, every other type of column scan still has to provide an implementation (currently just resorts to a unimplemented! call). Furthermore, every new type of ColumnScan has to be entered into the ColumnScanEnum object, which is tedious. A third issue is that in some cases, the type of the ColumnScan is known (e.g. in TrieProject one knows that the column scans are ColumnScanReorder), but they are still stored as a vector of ColumnScanTs. In order to limit the amount of "unpacking" one has to do, currently, all functions specific to one type of ColumnScan are also implemented on the ColumnScanEnum which panics if the function is called on the wrong type of column scan. However, it may be difficult to avoid this in general because sometimes we store several different types of ColumnScans in the same vector, e.g. in TrieJoin the scans could either be ColumnScanJoin or ColumnScanPass.

To solving those problems you would probably:

• Instead of ColumnScanEnum use dyn ColumnScan
• Have ColumnScanTypeT enums for each type of column scan

Implementing the dyn approach can also be done first without the rest for testing the performance impact. In this case one would leave the extra functions pos and narrow in the trait.

[matzemathics]

This crate might be of interest for cleaning up the glue code without compromising on performance: https://docs.rs/enum_dispatch/latest/enum_dispatch/