keep_dims for reducers

[wolfv]

While implementing std and var we need to use a reducer-in-a-reducer that is multiplied with something -- so keep_dims is necessary for successful broadcasting.

Since keep_dims would modify the return type of the reducer for fixed_shape and static dimensioned arguments, we need to have this as a template argument.

One idea would be to make the evaluation strategy somewhat composable. I am currently thinking of this syntax:

sum(a, keep_dims() | evaluation_strategy::immediate())

This would be deduced into an object that carries template arguments for evaluation strategy, keep dims (and potential other compile time args we want to supply).

I guess we could also overload the comma operator but that would be a bit crazy.