Manually calling an Initializer from Python

I'd like to initialize my LSTM with an orthogonal matrix, and the only way to do this currently is manually (#529). Of course I could create an orthogonal matrix in NumPy myself, but since this is already implemented in `SaxeInitializer`, it would be nice to be able to call its `initialize_params` method without passing it to a parameter's constructor. The C++ code uses `TensorTools::randomize_orthonormal` but that doesn't seem to be accessible from Python either.

Start by reading the Python-facing SaxeInitializer and its initialize_params method, then inspect the C++ TensorTools::randomize_orthonormal reference mentioned in the issue. The change is complete when an initializer can be called directly from Python for an LSTM without passing it to a parameter constructor, with the existing orthogonal initialization behavior preserved.

Written by the indexing model from the issue text.