Clarify floating-point sort order for ascending and descending sort

[kgryte]

Currently, the specification does not indicate the behavior for sorting the IEEE 754 floating-point values: NaN, -0, and +0.

In NumPy, NaN values are sorted to the end and signed zeros are not sorted:

In [1]: np.sort([2.0, 0.0, 1.0, -0.0, np.nan, np.nan, 3.0])                                                                                   
Out[1]: array([ 0., -0.,  1.,  2.,  3., nan, nan])

When sorting lists in Python, NaN values are left in place and signed zeros are not sorted:

In [1]: list = [2.0,0.0,1.0,-0.0,float('nan'),float('nan'),3.0,-1.0];                                                                         

In [2]: list.sort()                                                                                                                           

In [3]: list                                                                                                                                  
Out[3]: [-1.0, 0.0, -0.0, 1.0, 2.0, nan, nan, 3.0]

Other languages have made different design decisions. E.g., Julia sorts signed zeros and places NaN values at the end.

julia> sort([2.0,1.0,0.0,-0.0,NaN,NaN,3.0,-1.0])
8-element Array{Float64,1}:
  -1.0
  -0.0
   0.0
   1.0
   2.0
   3.0
 NaN  
 NaN

IMO, we should clarify the sort order for ascending and descending sort. Preferably,

• sorting signed zeros
• sorting NaN values to the end when sorting in ascending order and to the beginning when sorting in descending order

[kgryte]

By specifying sort order for floating-point values, this has implications for unique (see gh-249), as sorting NaN values to the ends affords relatively straightforward workarounds for handling multiple returned NaNs as discussed in the referenced issue.