Fix and improve map reduce

[AnesBenmerzoug]

Description

This PR closes #193 and closes #165

Changes

• Adds a SequentialParallelBackend class that runs all jobs in the current thread.
• Defines a BaseParallelBackend abstract class.
• Defines a AbstractNoPublicConstructor abstract metaclass to disallow directly instantiating parallel backends.
• Passes inputs to MapReduceJob at initialization.
• Removes chunkify_inputs parameter from MapReduceJob.
• Removes n_runs parameter from MapReduceJob.
• Calls the parallel backend's put() method for each generated chunk in _chunkify().
• Skips chunkification in MapReduceJob if n_runs >= n_jobs.
• Use singledispatch and singledispatchmethod to handle different input types in chunkify and _get_value.
• Adds a __repr__ method to ValuationResult.
• Calls parallel_backend.put() on the input to MapReduceJob's __call__ method if it is a sequence of Utility objects.
• Renames ParallelConfig's num_workers attribute to n_local_workers.
• Small fixes to MapReduceJob's docstring.

Checklist

• Wrote Unit tests (if necessary)
• Updated Documentation (if necessary)
• Updated Changelog
• If notebooks were added/changed, added boilerplate cells are tagged with "nbsphinx":"hidden"

[AnesBenmerzoug]

@mdbenito I have tried using partial to avoid using map_kwargs and reduce_kwargs but it didn't work with ray's put function. I checked their code and it does different things depending on the passed input.

[mdbenito]

This is looking nice :) I still have my previous question about putting everything to the backend, though. And a few comments here and there

[AnesBenmerzoug]

@mdbenito I have made the changes we talked about:

• removed n_runs from MapReduceJobs
• call put() on each chunk inside _chunkify()

I also added two tests for MapReduceJobs and fixed a bug when passing numpy arrays to _chunkify().

[mdbenito]

I think this looks rather nice now :) I suggested a couple of cosmetic changes, ditching redundant elses, but other than that I'm 100% fine with merging. Go ahead!

[mdbenito]

Suggested change

	if map_kwargs is None:
	self.map_kwargs = dict()
	else:
	self.map_kwargs = {
	k: self.parallel_backend.put(v) for k, v in map_kwargs.items()
	}
	self.map_kwargs = {
	k: self.parallel_backend.put(v) for k, v in (map_kwargs or {}).items()
	}

[AnesBenmerzoug]

I think this makes it harder to read since this is not a common pattern.

[mdbenito]

Suggested change

	if reduce_kwargs is None:
	self.reduce_kwargs = dict()
	else:
	self.reduce_kwargs = {
	k: self.parallel_backend.put(v) for k, v in reduce_kwargs.items()
	}
	self.reduce_kwargs = {
	k: self.parallel_backend.put(v) for k, v in (reduce_kwargs or {}).items()
	}

[mdbenito]

Why don't we always directly import like this instead of using the if TYPING guard and then quoting all types? (that's a bit of a PITA, tbh)

[AnesBenmerzoug]

I think it was a leftover from when we didn't pin the minimum numpy version to 1.20.

The typing module was added to numpy in version 1.20.

So, yes I think we should just import it directly.