Msaf

-------

* Compatibility with Librosa 0.6.0 when computing CQTs.
* Added VMO algorithm.
* Modified quick-guide with clearer example for evaluation.

-------

* Fixed hierarchical boundaries, which didn't work on JAMS 0.3.0.

-------

* JAMS bug supposedly fixed in v0.1.5 was not really fixed. Fixed now. Alrighten.

------

* Added `framesync` parameter to `compute_features.py` script example
* `joblib` was repeated in the requirements. Fixed
* Updated PyPi docs strings such that it shows that MSAF if Python 3.56 compatible
* Making `configparser` compatible with Python 2.7
* Using latest `scluster` method, from McFee's code [here](https://github.com/bmcfee/lsd_viz)
* Adopted code to JAMS 0.3.0.

------

* Fixed bug tha threw a `TypeError` if multiple algorithms were run in a single JAMS file with `None` and other label_ids in it
* Added new `vmo` oracle segmentation method (by Cheng-i Wang, thanks!)
* Adapting sonify function to latest numpy
* Using KMeans from sklearn instead of scipy for 2D-FMC. Results are better
* Making sure we are never using more number of clusters than number of segments for 2D-FMC
* Added new parameter `2dfmc_offset` in the 2D-FMC method
* Using np.inf normalization for 2D-FMC now, since it seems to yield better results (at least for Beatles TUT)
* Padding beat-sync features now, seems to fix potential misalignment of boundaries. Some algorithms (2D-FMC, CNMF) seem to yield better results now
* Modified features file: two new fields may be addeed: `est_beatsync_times` and `ann_beatsync_times`.
* The member variable `_framesync_times` in the `Features` was never updated. Fixed it

------

* Included Python 3.5 in the metadata
* Removed old functions from i/o module that nobody should be using
* cleanued up code for reading/writing estimations (just a tiny bit)
* Unit tested i/o module