Option to serialize a source

[kyleaoman]

It's probably reasonable to expect a user to be able to hold (1) the particle data of the source and (2) the datacube in memory simultaneously as a minimum memory requirement. As currently implemented, however, martini is more memory-hungry than this because it pre-computes and stores the spectra of all particles simultaneously, requiring nchannels * nparticles floating point values, which for some setups could be a very large number. Pre-computing the spectra is done for speed, but it would be possible to cut down on memory by segmenting the source into a series of chunks of ~equal numbers of particles. The simplest implementation would be to run the source insertion loop once for each such chunk, which is a significant hit in terms of cpu time cost. Giving some flexibility to keep memory requirements under control would be good though. Could explore other options like optionally using a memmap to store the spectra on disk rather than in memory. Costs a bit of access speed but maybe better than running the main loop more than once.