Benchmark differences after dask-expr switch

[phofl]

• test_filter: Mostly because of the missing arrow conversion at the moment

• test_cow_chain: same, will re-check when we have merged the convert string option for dask-expr

• test_join_big_small very likely the same reason

• test_q6 in test_h2o: we are fusing the input parts based on the number of columns which reduces the partition count. We are not utilising the cluster completely since we don't have enough partitions, but this should not really matter if you have larger datasets, so I am inclined to ignore this

[phofl]

Runtimes are now where I would expect them to be, a little bit faster but not much