faster sorting for -fastxtile- or -by- and -tab- operations

[laszlosandor]

Without digging into it, the findings from the Statalist thread could help binscatter speed up on big data: http://www.statalist.org/forums/forum/general-stata-discussion/mata/172131-big-data-recalling-previous-sort-orders

[laszlosandor]

@michaelstepner: This is still a shame that this works, but it looks like it does. Implying an order (calculable for categorical variables, e.g.) is still faster than a generic -sort-.

[laszlosandor]

Maybe this is more useful: Even if one does not play with a faster sort (though imposing an order calculable from bin-counts should be much faster than sorting), at least we could avoid sorting part of the data never used:

sort `touse'
count if `touse'
sort `touse' sortvar in `touse_first'/`touse_last'

That said, the saving need not be game-changing, as x_r is . when touse is 0, so sorting will be trivial. Might matter slightly more for sorting on by_var or x_q.