Kernel centres semi-stratified selection without replacement

[mierzejk]

The pull request revises the way kernel centres are selected. The following changes have been introduced:

• np.random.choice to sample input data without replacement;
• np.percentile so samples are stratified with respect to (possibly multivariate) x values.

I call this manner of centre selection semi-stratified, because the final result is concatenated independently from every column of the second array dimension, where indices are chosen randomly from quantiles returned by np.percentile.

Resolves the following issues:

• fixes #12 - sample without replacement.
• fixes #13 - increases result stability, as confirmed by the original poster.

[mierzejk]

Please note that all commits covered by this pull request are also included in #23 Aggregated pull request.