Move distributed pixel data to node shared memory

[tskisner]

Currently, the PixelDistribution and PixelData classes store a sparse set of pixel values on each process. These pixel values correspond to the pixels covered by the local detector data on each process. For some scenarios (high resolution pixelization, high cadence scanning), the sky fraction covered by a single process can be quite large. The impact of this high per-process memory use is that fewer processes per node can be used.

One possible solution is currently in development which uses MPI shared memory for storing a node-local copy of pixels seen by all processes on the node. This incurs a cost when updating the pixel values, since locking or similar turn-taking must be done by processes on the node. However, the hope is that this cost will be more than offset by enabling the use of more processes per node.