Reconstructing large volumes

[talonchandler]

This WIP implements two strategies to scale our reconstructions.

1. stretched_multiply enables efficient application of a small OTF to a large dataset. This approach approximates the OTF as a piece-wise constant function, and applies this smaller array to datasets without upsampling. This approach allows us to reduce OTF memory requirements by as much as needed, trading OTF memory for accuracy.

2. An overlap-add approach to large convolutions. This approach avoids larger-than-GPU-memory FFTs by splitting the volume in to blocks, processing each block individually, then combining the results. Here I am implementing a "for loop" approach to the outer blocks (so it will scale linearly), but when we integrate with recorder we can parallelize the processing of each block on the HPC.