Version 4.0

1. Added quantized models to all Malaya models, reduce inference time by 2x and model size by 4x.
2. Retrain constituency parsing, improved accuracy slightly by ~1-2%.
3. Added vectorization interface for sentence / word level for all classification models.