Automatic python bindings

[rcurtin]

I've spent a long time working on this and there is still a lot to do before merge, but they should at least be in a usable state and can be played with now.

There are some major changes in this PR:

• Instead of mlpack::util::Option being used with CLI::Add(), now Option is a template typedef specified in <mlpack/core/mlpack_main.hpp> and controlled by the MLPACK_BINDING_TYPE macro.

• CLI has been revamped (and should probably be renamed) and now does not actually do the parsing of the command line. It simply is a singleton that holds options and their current values, and it also handles the timers too.

• Each option that CLI holds has a type, but since we can't deduce the type from a boost::any, now we use this CLI::functionMap structure along with the C++ typeid to call the appropriate functions for a given option. Before calling CLI::Add(), an option is responsible for adding all of the function mappings that will be used by that program/binding type.

• src/mlpack/bindings/cli/ is a reimplementation of the command-line program, but with the refactored CLI code. Here, CLIOption defines the functions that we need for command-line programs.

• src/mlpack/bindings/python/ contains automatically-generated Python bindings controlled by the add_python_binding() CMake macro.

If you have cython installed, you can check this branch out and build it, and then you can actually use the python bindings:

$ pwd
/home/ryan/src/mlpack-rc/build/
$ cd src/mlpack/python/bindings/build/lib.linux-x86_64-2.7/
$ export LD_LIBRARY_PATH=/home/ryan/src/mlpack-rc/build/
$ export PYTHONPATH=.
$ python
Python 2.7.13 (default, Dec 18 2016, 20:19:42) 
[GCC 6.2.1 20161215] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from mlpack import pca
>>> import numpy as np
>>> x = np.random.rand(5, 5)
>>> x
array([[ 0.1659253 ,  0.44932094,  0.89157297,  0.94657701,  0.84910244],
       [ 0.99556297,  0.15514302,  0.52927208,  0.52319716,  0.73886948],
       [ 0.80643328,  0.04821054,  0.97173154,  0.65464406,  0.48690407],
       [ 0.29296537,  0.7519181 ,  0.33787851,  0.72429684,  0.13147376],
       [ 0.01446388,  0.04444607,  0.99994722,  0.33920724,  0.46390401]])
>>> y = pca.pca(input=x, new_dimensionality=2)
>>> y
{'output': array([[ 0.23981584, -0.20722268],
       [-0.53626128,  0.30452866],
       [-0.4147487 , -0.05555313],
       [ 0.55255013,  0.4990668 ],
       [ 0.15864401, -0.54081965]])}
>>> 

There are still many things to do before merging this in though, although maybe some of them can be separated into separate tickets to be resolved after merge:

• Update CMake configuration so that Python isn't required for the build.
• Handle Python binding generation via a CMake variable.
• Install Python bindings to the appropriate place on make install.
• Test the components of the Python bindings automatically, instead of by hand.
• Test the components of the CLI bindings automatically, instead of by hand.
• Figure out a way to write tests for the command-line programs, which might require some kind of DSL, I'm not sure yet.
• Handle DatasetInfo/matrix objects in the Python bindings---right now they are ignored. For this I need to see what the "typical" way is of describing what your dimensions are.
• Update PROGRAM_INFO strings to be correct for both CLI and Python bindings.
• Check that these bindings can work with pandas dataframe objects and other commonly-used Python data science tools (I'm pretty sure they don't, so I'll need to fix this).

[rcurtin]

Ok, I think this is ready to merge; I've been sitting on it for quite a while but any outstanding issues have been fixed. I'll let this sit for another 3 days in case anyone has any comments.

[zoq]

Sounds good, can't wait to work with the bindings.