Read a reasonable amount of data per chunk.

In read_csv (and others), we are reading only 100 rows or so at a time. This gives kb-sized chunks, the reader is likely optimal in the mb-sized ranges.