Changed the code structure into more efficient way to separate graphs building of prediction process

[OmarSayedMostafa]

the following are the main updates:

• created the ClinicalConceptExtraction class who contains the elmo model graph and clinical concept graph built and loaded in initiate time to avoid building them with every prediction process.

• edited the build input method to work in a direct way by get embeddings and labels by the same batch size in one forward pass, instead of looping over sentences and get embeddings for each sentence and use tf.dataset to create padded iterator by the given batch size.