Fastfm

This release adds the capabilities to build binaries (wheels) that work across a large range of Linux distributions to our travis-ci. This will greatly benefit everyone who had issues with source installing fastFM.

Special thanks goes to

- zouzias for pointing us to the manylinux project,
- to chittti for writing the manylinux based build script
- and chezou for integrating the build script in our travis config.

This release contains again various community contributions :smile: .

* The default blas version has been changed from cblas to openblase (thanks iramykytyn),
* python 3.6 support has been improved (thanks chezou),
* warm start has been added to als classification (thanks lucidfrontier45) and
* changes in the numpy (thanks mkdy) and sklearn api (thanks takuti) have been incorporated.

This release fixes pip install for ubuntu 16.04 82 and all other systems that have a pip version > 7.0.

This release fixes a bug leading to segmentation fault errors for the BPR solver.
The bug was reported (by felixmaximilian) in 61 and fixed by https://github.com/ibayer/fastFM-core/pull/13 .

The release further improves code quality by fixing a fair amount of coding style violations (thanks to takuti ).

diff
- fastFM has been accepted as contribution to the JMLR OSS track.


[JMLR OSS track](http://www.jmlr.org/mloss/)

fastFM supports now whatever BLAS version is installed (OpenBLAS is prefered).

This minor release updates the fastFM-core sub-module which contains a fix for 37 .
Previously the same random number generator seed was used for each start of the mcmc chain, which lead to bad mixing with warm start ( `n_more_iter` parameter). This issue is now fixed by using a new random seed for each warm start.