Better test column ops

As #1337 suggests, our column ops are not sufficiently tested. In particular, it's not enough to test packed/non-packed versions, but, for the packed case, we also need to test both `NUM_INT_PACKS==NUM_MID_PACKS` as well as `NUM_INT_PACKS>NUM_MID_PACKS`. Given pack size restrictions, if num_mid_levels is a power of 2, then we always fall in the second case (provided pack size is <= num_mid_levels). That's the only scenario we were testing (and it's the one interesting for scream, which will operate at 128 levels). But some tests run with less levels (e.g., 72 levels for our homme-relatd tests), and for some pack sizes (e.g., the default, 16) can fall in the first scenario, which in some cases can trigger a different impl in ColumnOps.

Hence, we need to test something like `num_levels=n*SCREAM_PACK_SIZE` as well as `num_levels=n*SCREAM_PACK_SIZE + SCREAM_PACK_SIZE/2`.