This checks that buildings which cross tile boundaries are more likely to be assigned the correct landuse kind. The issue was that, before the change from query-per-layer to query-per-table, we were clipping the landuse to the tile but the building to a 3x3 tile area. This meant that large buildings had much less overlap with landuse polygons, and some would not be assigned a `landuse_kind`.

Now we clip all inputs (except water) to the tile, which solves this problem, but means that the same building in adjacent tiles could be assigned different `landuse_kind`s.

Additionally, it looks like `kind: building_part`s weren't being assigned a `landuse_kind`, although the test now passes, so something we changed has fixed that too.