Improve parallel testing

We have hacked together a testing infrastructure involving running `nosetests` or `python -m unittest discover` at the top level (skipping all tests that require `mpiexec`) and then running `mpiexec -n nosetests` in various subdirectories to run some tests using MPI. It would be nicer if we could run `nosetests` at the top level and discover and run all the tests, spawning MPI processes as needed, and printing non-garbled results.

Something like this: https://github.com/dagss/mpinoseutils.git