NBytes estimates on pandas dataframes with strings are overly pessimistic

This results in Dask swapping to disk needlessly. Now that we're better at checking process-level memory we might consider relaxing our spill-to-disk mechanism based on estimated byte costs on a per-data-element basis.