PERF: using _reduce_axis1 to improve perf in df.sum,min, max with axis=1

[jbrockmendel]

DataFrame reductions with axis=1 can be expensive when there is more than 1 backing array. _reduce_axis1 improves on that for any/all. We should be able to do the same thing for sum, prod, min, and max, though it'll take more effort.

Might be a good fourth issue, not first.