Make the models fine-tunable

[davidalami]

Current models cannot be fine-tuned because we do not have the code of Keras layers inplace. I was lucky enough to find the code from original authors of ConveRT. Assistance required to adapt this code into the current repository.

[lhr0909]

I could potentially take a look at this soon. I am currently working on a service wrapper around this module for getting the sentence features from ConveRT. I would like to be able to fine-tune the model and use it for English text understanding and response selection tasks.

[davidalami]

Hey Simon, excellent! Thank you for your enthusiasm, assigning the task to you :)