Switch mp to joblib in asPointwise

[JesseCresswell]

This PR addresses issue #19
I used this test code to demonstrate the issue and benchmark the solution:

import numpy as np
import skdim
import time

np.random.seed(0)
X = np.random.random((20000, 10))

estimator = skdim.id.ESS() # LocalEstimator
n_jobs = 1

t0 = time.time()
lid = skdim.asPointwise(X, estimator, n_neighbors=10, n_jobs=n_jobs)
t1 = time.time()

print(lid.mean())
print(t1-t0)

On master, commit b9e8845, by only changing n_jobs I observe the following using a computer with 40 cpu. Note that the timing includes knn computation which is affected by n_jobs, but for these settings knn takes only a couple seconds for n_jobs=1.

n_jobs	Time (s)	Mean LID	Notes
1	222.62	8.745108	About 50% of cpu being used at 100% (others at 0%)
2	-	-	Does not complete. Makes no progress. 100% of cpus at 0% usage

After the PR changes I observe

n_jobs	Time (s)	Mean LID	Notes
1	219.36	8.745108	About 50% of cpu being used at 90%
2	208.17	8.745108	About 100% of cpu being used at 90%
20	28.47	8.745108	About 100% of cpu being used at 100%
40	16.95	8.745108	About 100% of cpu being used at 100%

Conclusions:
Multiprocessing is now working properly and there is a significant speedup from increasing n_jobs (including both knn computation, and estimator computation). The computed LID values are unchanged.
The ESS estimator code apparently uses more than one cpu even when n_jobs=1 , although it is not doing so efficiently. I have not tried to isolate which step is trying to use multiple cpus.

[j-bac]

thanks! Looks good. Just a remark that ESS().fit already applies the estimator pointwise since it is a LocalEstimator so this example should not be used in practice

[JesseCresswell]

Yes fair point for anyone else reading this. The built-in method ESS.fit() appears to be more efficient than using asPointwise. My example was just for benchmarking.