Single-machine, multi-process integration

This MR adds the MultiprocIntegrator class for distributing the work over a set of processes on the local machine.

To simplify task distribution, each worker (process) handles one full dataset (finer granularity would involve IPC).

To do

• Refactor the classes to use a base class DistributedIntegrator
• Ensure that distributed integration (over SLURM) still works
• Test MultiprocIntegrator
• Automatic resources handling:
  • Determine automatically the best affinity mask
  • Pick the best opencl platform
• Plug MultiprocIntegrator to the CLI/configuration file
• Clean-up the test/debug code fragments (eg. os.environ, is_power9, etc)