Drop is translated into a Projection which creates issues for inconsistent partitions

see ``test_drop_meta_mismatch`` in dask/dask

drop currently is translated into a Projection which is based on meta, so if some partition don't have some of those columns then you will run into KeyErrors, this can potentially be problematic.

A potential workaround is to create it's own Expression class and let it inherit from Projection so that it is taken into account in all optimisations, I have to think a little bit more about this if this is really a good idea

Start with test_drop_meta_mismatch in dask/dask and trace how drop is translated into a Projection. Investigate how inconsistent partition columns lead to KeyErrors, then evaluate the issue's proposed Expression subclass approach while preserving optimization behavior; done means the test passes without errors for mismatched partition metadata.

Written by the indexing model from the issue text.