`scheduler` kwarg for `mapreduce` based functions

I think it'd be useful if we added a keyword argument to `mapreduce` and `mapreduce` based functions that lets one swap out the reduction implementation. Then a package like Distributed, OhMyThreads, CUDA, etc can define and export schedulers which can be passed to `mapreduce` to change the details of the scheduling (i.e. multithreading, multiprocessing,  GPU, etc)