This PR closes #600.

The various SKLL modules learner.py, experiments.py, config.py etc. were getting really long as single files and unwieldy when it came to adding new development. This PR tries to address this issue by converting these modules into sub-packages instead with specific functions moved to different modules under these sub-package.

Here's what the SKLL code tree looks like today:

.
├── __init__.py
├── config.py
├── data
│   ├── __init__.py
│   ├── dict_vectorizer.py
│   ├── featureset.py
│   ├── readers.py
│   ├── writers.py
├── experiments.py
├── learner.py
├── logutils.py
├── metrics.py
├── utilities
│   ├── __init__.py
│   ├── compute_eval_from_predictions.py
│   ├── filter_features.py
│   ├── generate_predictions.py
│   ├── join_features.py
│   ├── plot_learning_curves.py
│   ├── print_model_weights.py
│   ├── run_experiment.py
│   ├── skll_convert.py
│   └── summarize_results.py
└── version.py

13 directories, 84 files

With this PR, it will look like:

.
├── __init__.py
├── config
│   ├── __init__.py
│   └── utils.py
├── data
│   ├── __init__.py
│   ├── dict_vectorizer.py
│   ├── featureset.py
│   ├── readers.py
│   ├── writers.py
├── experiments
│   ├── __init__.py
│   ├── input.py
│   ├── output.py
│   └── utils.py
├── learner
│   ├── __init__.py
│   └── utils.py
├── metrics.py
├── utils
│   ├── __init__.py
│   ├── commandline
│   │   ├── __init__.py
│   │   ├── compute_eval_from_predictions.py
│   │   ├── filter_features.py
│   │   ├── generate_predictions.py
│   │   ├── join_features.py
│   │   ├── plot_learning_curves.py
│   │   ├── print_model_weights.py
│   │   ├── run_experiment.py
│   │   ├── skll_convert.py
│   │   └── summarize_results.py
│   ├── constants.py
│   └── logging.py
└── version.py

12 directories, 91 files

In addition, some functions that were previously indicated to be private (with a leading underscore) are now public since they can actually be quite useful as part of the API. Some examples include experiments.input.load_featureset() and experiments.output.generate_learning_curve_plots().

Specifically, this PR:

• Reorganizes and refactors SKLL module code into sub-packages.
• Sets up __init__.py for these sub-packages such that most of the idiomatic SKLL imports work just like they do now.
• Removes unnecessary imports from the top-level skll namespace. Right now only the following are importable from skll - FeatureSet, Learner, and run_configuration.
• Moves logutils.py to utils/logging.py
• Consolidates all the constants from across the various modules into a single place: utils/constants.py.
• Moves all the command line utilities under utils/commandline/ and adjusts setup.py accordingly.
• Converts the private _import_custom_learner() function to a public function called load_custom_learner() which does not modify globals() directly but just returns the loaded learner to the caller which then modifies globals() on its end.
• Renames test_logutils.py and test_utilities.py to test_logging_utils.py and test_commandline_utils.py.
• Updates API documentation to remove things that are not top-level importable, to include the new sub-packages and constants, and to tweak the names of sections.