Parallelise with ray

[Xuzzo]

• New interruptible worker with ray.
• All Shapley methods are now independent of map_reduce.
• truncated_montecarlo_shapley is the new default method
• adds tests for truncated_montecarlo_shapley

Closes #22

[Xuzzo]

Sorry for the deleted type hints. Upon merging I had a bunch of conflicts and in the attempt to solve them I removed some of your stuff. Will fix it now.

[AnesBenmerzoug]

One change that we could make later if we really want to have an abstract interface for the parallelization would be to switch from decorating the classes with @ray.remote to applying it in the __new__ method. i.e. change from

@ray.remote
class ShapleyWorker:
   ...

to for example:

class RayWorkerBase:
   def __new__(cls, *args, **kwargs):
       remote_cls = ray.remote(*args, **kwargs)(cls)
       actor = remote_cls.remote()
       return actor

class ShapleyWorker(RayWorkerBase):
   ...

Of course, this is just an idea

[Xuzzo]

I tried your suggestion (the one about RayWorkerBase), but there are some things not working properly. Will keep the decorator for now and leave the abstract interface for the future. Also, I have removed the fake remote method from the worker and coordinator and just ignored the type since it was giving me annoying warnings.

[AnesBenmerzoug]

Looks good to me. Feel free to merge.