Parallel

[rphes]

Parallel implementations of K-Modes and K-Prototypes. See #76 for more information.

The approach is mostly identical to scikit-learn's implementation of parallel K-Means. Changes are:

• allowing multi-core reproducibility by propagating random seeds to all jobs
• extracting the parallel unit of work, a single run, to a separate function
• performing units of work sequentially if n_jobs=1, or using joblib otherwise

Note that it only makes sense to use parallel execution for n_init>1. This is not enforced in code.

I updated the documentation and tests accordingly. The expected outcomes of K-Modes tests had to be changed due to the new method of seeding.

When this PR is merged, I will look into updating the examples/benchmarks to make use of the changes.

[coveralls]

Coverage increased (+0.1%) to 96.706% when pulling e2ca4b1 on rphes:parallel into 73b4ffe on nicodv:master.

[nicodv]

Beautiful work, @rphes !