This avoids the excessive memory requirements described in #77 by exploiting numpy views. Instead of allocating and zeroing a memory block equivalent to the final result, we now only use one timestep of the final result, creating the larger block only as a view into this much smaller allocation.

closes #48