Broadcast FourierKernel._convolve_volume

[garrettwrong]

During #406, it was noted that it should be straightforward to broadcast _convolve_volume instead of iterating. Right now it is almost generalized.

We'll just need to be careful about memory usage... Can consider a parameterized batch size, maybe defaulting to something based on voxels*volumes. Should double check whether the speedup is worth it first.

In the case it isn't, the code could be cleaned up (rm singleton branches etc).