Implement python wrappers for predictions and tune speed

[Bowid]

1. Predict now use std::set instead priority_queue. From my point of
   view it is better to maintain set of fixed size and drop predictions
   with too low score immideately instead storing them in ordered manner.
2. Implemented extension class of StarSpace to make things a bit more
   pythonic. It is definitely better to have return value instead of
   passing parameters to fill to functions. From my point of view, speed is
   not changed in case modern C++ compilers handle return of complex values
   properly without copying them.

[Bowid]

Also here is the sample, let it be here

import starwrap

model_path = '/path/to/trained/model'

args = starwrap.args()
model = starwrap.starSpace(args)
model.initFromSavedModel(model_path)

line = input("Enter text for classification: ")
parsed = model.parseDoc(line, ' ')

# Output is list of tuples in format (probability, token_id)
# Predict takes two arguments: parsed doc and maximum predictions count
tokens = model.predict(parsed, 10)

probabilities = [item[0] for item in tokens]
# Token_id is integer, thus it should be rendered to feature (string)
features = model.renderTokens(tokens)

for feature, probability in zip(features, probabilities):
    print('Feature {f} probability {p}'.format(f=feature, p=probability))