Skll

- Fixed summary output for ablation experiments. Previously summary files
would not include all results.
- Added ablation unit tests.
- Fixed issue with generating PDF documentation.

- Added two new _required_ fields to the configuration file format under the
`General` heading: `experiment_name` and `task`. See the
`run_experiment documentation <http://skll.readthedocs.org/en/latest/run_experiment.html#creating-configuration-files>`__
for details.
- Fixed an issue where the "loading..." message was never being printed when
loading data files.
- Fixed a bug where keyword arguments were being ignored for metrics when
calculating final scores for a tuned model. This means that **previous**
**reported results may be wrong for tuning metrics that use keywords**
**arguments**: `f1_score_micro`, `f1_score_macro`,
`linear_weighted_kappa`, and `quadratic_weighted_kappa`.
- Now try to convert IDs to floats if they look like them to save
memory for very large files.
- `kappa` now supports negative ratings.
- Fixed a crash when specifing `grid_search_jobs` and pre-specified folds.

- Hotfix to fix issue where `grid_search_jobs` setting was being overriden
by `grid_search_folds`.

- Added `skll.data.write_feature_file` (also available as
`skll.write_feature_file`) to simplify outputting .jsonlines, .megam, and
.tsv files.
- Added more unit tests for handling .megam and .tsv files.
- Fixed a bug that caused a crash when using gridmap.
- `grid_search_jobs` now sets both `n_jobs` and `pre_dispatch` for
`GridSearchCV` under the hood. This prevents a potential memory issue when
dealing with large datasets and learners that cannot handle sparse data.
- Changed logging format when using `run_experiment` to be a little more
readable.

- Fixed serious issue where merging feature sets was not working correctly.
**All experiments conducted using feature set merging** (i.e., where you
specified a list of feature files and had them merged into one set for
training/testing) **should be considered invalid**. In general, your
results should previously have been poor and now should be much better.
- Added more verbose regression output including descriptive statistics
and Pearson correlation.

- Fixed all known remaining compatibility issues with Python 3.
- Fixed bug in `skll.metrics.kappa` which would raise an exception if full
range of ratings was not seen in both `y_true` and `y_pred`. Also added a
unit test to prevent future regressions.
- Added missing configuration file that would cause a unit test to fail.
- Slightly refactored `skll.Learner._create_estimator` to make it a lot
simpler to add new learners/estimators in the future.
- Fixed a bug in handling of sparse matrices that would cause a crash if
the number of features in the training and the test set were not the same.
Also added a corresponding unit test to prevent future regressions.
- We now require the backported configparser module for Python 2.7 to make
maintaining compatibility with both 2.x and 3.x a lot easier.