Vectorise the population decoder in numpy

[celefthe]

Closes #204.

process decode fits the population logistic regression and linear discriminant in numpy across windows, folds and shuffles at once, as it already did for the one-feature decoders, instead of one scikit-learn fit per fold.

• decoding.population_decisions, with population_lda_decisions (scikit-learn's lsqr solver with Ledoit-Wolf shrinkage) and population_logistic_decisions (batched Newton steps with backtracking).
• region_decisions fits the one-feature logistic regressions through population_decisions, and logistic_decisions is removed.
• decoding_table and rolling_table take the population scores, nulls and weights from population_decisions. fit_scores stays as the scikit-learn reference in the tests.
• Population logistic scores and weights change slightly, since the Newton fits converge further than scikit-learn's default tolerance.
• Tests in tests/test_process.py, docs/how-to/decoding.md and a changelog entry.