spawn method context

[erikhuck]

The default spawn method is fork which apparently duplicates all the memory of the main process into the child. If gpu-tracker does that by default on linux machines, that creates a problem for jobs ran on those machines. Instead, we can have the tracker process use its own context that uses the 'spawn' method instead of 'fork' like so:

import multiprocessing as mp

ctx = mp.get_context('spawn')
class _TrackingProcess(ctx.Process)

If this works as expected, we should test out how memory usage differs when using the 'spawn' method versus 'fork'.

• Create a script that first allocates a bunch of memory.
• After all that memory is allocated, create a Tracker object and start tracking.
• See how memory differs using the 'spawn' method versus the 'fork' method.
  We probably want to do the above test using the global setting with mp.set_start_method to see if there truly is an issue before we bother with using get_context and inheriting from ctx.Process.