Binary dump bunch function

[austin-hoover]

A while ago we mentioned a possible bunch.dumpBunchBinary function that will write the coordinates to a binary file instead of text file. This would avoid enormous files for large bunches that result from bunch.dumpBunch. Similar to numpy.savetxt vs. numpy.save.

[azukov]

Possible ideas

1. Inject some functionality into Bunch.dump(), the MPI code will stay the same, but disk writing pieces will change
   • Save as numpy using cnpy library or similar
   • Use HDF5 dependency and probably HighFive to save in H5
2. Implement Bunch.collect() that will move Bunch to RANK0 and then make Bunch.to_numpy() that will return some numpy array.

[austin-hoover]

I'm guessing HDF5 would be better that NumPy? There is other information in the current dumped files besides the phase space coordinates, like synchronous particle parameters. There are also optional particle attributes that can be saved, and some nodes like tune calculators can add columns to the file. I'm not sure if NumPy can save this additional info. Currently it's all in the comment lines of the file.

[azukov]

@woodtp @ruisardkj We can also start with pure python implementation, it will uses MPI and will ask all nodes to send batch of particles (10k) and then convert values into numpy. This way we can potentially have bunch that doesn't fit into one node's memory if we use disk backed numpy instance.

[woodtp]

@azukov I've just created a draft PR #72 as a first pass at a pure python implementation. So, far I wrote a function which takes info from the Bunch and loads it into a dict of numpy objects. I haven't incorporated any sort of batching yet, so you could run out of memory if you make a big enough bunch.