Perform staging in the scheduler

Currently, operations on `DArray`s are staged into thunks before being passed into the scheduler, removing all information about operation structure. For some operations, such as a matmul, knowing the original structure of a set of operations might allow us to substitute an optimized implementation or hardware-specific communication pattern, even calling out to an external library (like Elemental). Users might also want to provide their own operations, and together with #166, may want to provide more efficient ways to compute them. In short, we should move operation staging into the scheduler, and when #166 is implemented, we can pass the raw operation to the user's scheduler implementation and let it choose whether to stage to thunks, or to do something special.