Reference Issues/PRs

Towards scikit-learn#23958

What does this implement/fix? Explain your changes.

This simplifies the original implementation of PairwiseDistance by @jjerphan, with the following differences:

• PairwiseDistance{32,64} doesn't subclass BaseBaseDistancesReduction{32,64} anymore.
• This allows to add _parallel_on_{X,Y} methods to PairwiseDistance{32,64}, since these methods are decorated with @final in BaseBaseDistancesReduction{32,64} and thus can't be overwritten.
• This also remove the chunk computing mechanism, by considering only the case chunk_size = 1, as proposed by @ogrisel in this comment.
• This doesn't implement the Euclidean specialization yet to make benchmarks simpler.

Following this benchmark, we found that this PR yields a significant performance regression when n_jobs = 1 and an improvement when n_jobs > 1, for both euclidean and manhattan distances:

Any other comments?

As suggested by @jjerphan, decorating DistanceMetric subclasses with @final could alleviate some of the overhead and make this implementation competitive with main when n_jobs=1.