[C++] have GradientDescentParameters track a num_steps_averaged field

[suntzu86]

BLOCKING: #390

Before we can implement polyak-ruppert averaging for C++'s gradient descent function, we need the parameter struct to hold this value.

Currently, the moe.optimal_learning.python.cpp_wrappers.optimization.GradientDescentParameters wrapper takes in a num_steps_averaged field but does not pass it on to C++. The field is stored locally in Python so the data is not lost.

We need to:

1. add num_steps_averaged to the C++ GradientDescentParameters struct
2. update docs to describe averaging
3. remove the local copy stored in the Python wrapper object