Nltools

*New Functionality*
- Added ability to perform hyperalignment using procrustes transform and shared response model.
- Added ability to decompose across space and time
- Added cosine similarity to Brain_Data.similarity

*Bug Fixes*
- fixed bug with thresholding
- fixed bug with random seed and parallel processing bootstrap and permutations
- fixed bug with naming in design_matrix
- fixed bug with file_reader
- fixed bug with plotting (updated matplotlib and nilearn version)
- moved external code to /external folder

**New Functionality**
Added ability to decompose Brain_Data objects using PCA, ICA, NNMF.

**Bug Fixes**
Fixed bug with pypi with requirements file.

This release was lead by ejolly and added a bunch of new things.

**New Functionality**
- Added a new Preference class to specify preferences for nltools. The most notable one is which brain mask to implicitly use. For large datasets, 3 x 3 x 3 will dramatically reduce RAM costs and speed up computations. These masks are packaged in resources folder as well as versions that don't include the ventricles.
- ARMA estimator for regress methods
- robust regression for regress methods

**API Changes**
- Brain_Data.threshold has a new api. Can now accommodate bidirectional thresholding, but must specify an `upper` and `lower` bound thresholds

**Bug Fixes**
- small bug in bootstrap function
- requirements file now synced with setuptools
- rolled back matplotlib to avoid a new plotting bug.
- fixed data type output for Brain_Data.ttest()
- fixed an indexing bug with stats.downsample function

**Bugs Fixed**
- Brain_Data slicing
- optional imports
- refactored data module

**New Functionality**
- permutation and bootstraps now run in parallel
- can now concatenate Brain_Data and Adjacency objects

Fixed bugs with:
-ROC (forced choice and balanced accuracy, store gaussian weights)
-create_sphere
-Brain_Data logical indexing
-compatible with scipy 1.0.0
-winsorize

Added new functionality:
-transform_pairwise()

Now compatible with Python3