use colordering(DFPerm) and getindex(DFPerm) to squeeze multiple lt()
methods into one

- don't encode the indexing columns, use DataFrameRow hashes instead
- do only the parts of left-right rows matching that are required for a
  particular join kind
- avoid vcat() that is very slow for PooledDataVector
- now join respects left-frame order for all join kinds, so the
  tests/data.jl test were updated

sorting order is changed from NA first to NA last (it matches
the default data frame sorting)

so that DataFrameRow object doesn't need to be created

instead of using Dict{DataFrameRow,Int}, implement its own that
 - doesn't require DataFrameRow objects
 - calculates hashes in memory-efficient manner
 - keeps row hashes for efficient comparison

use it for join(), groupby(), nonunique()

disable DataFrameRowTests that use _RowGroupDict methods no longer
available

it's not used and it is no loner faster than nonunique()

by default no sorting is applied to preserve original ordering
(the initial order of the 1st rows is preserved) and make things faster