i believe i figured out how percentile is computed. this does
something very similar to median (for obvious reasons). at highlevel:
- get the index the percentile would land on and assume it falls
between indices
- take the values of the indices it falls between and figure out
weight of each value.
- handle the percentiles that fall on an exact index.

this is ~30x better. it's actually a bit slower (3%-5%?) than the
median computation so it's debatable if we want to remove median code.

drop scipy. also remove numpy.lib.recfunctions as it doesn't seem to
used anywhere.