high performance io

[jreback]

examine approach & integrations for the IO subsystems. pay attention to copies, memory mapping, and parallelization. let this be a master issue, with specific discussions in separate issues as needed.

• csv Much faster to_csv implementation (in libpandas) #34
• HDF5
• feather / parquet

in-line support parallel read / write (utilize dask)