restructure how models are loaded for core-slu-service.

[greed2411]

this requires discussion on how things models are going to be loaded in the future,

especially

1. presently a single model/workflow is being loaded (predict_wrapper in endpoints.py) on startup via config.yaml .this is even before accessing active configs from builder backend via an API. this isn't aligning with fetching configs from builder and then create inference functions for each of those models. Therefore we need to have different PREDICT_API for each of the CLIENT_CONFIGS, and start rewriting for that way.
2. continuing 1, means we need to load models only after receiving/creation of config from builder-backend. implies, no models will be loaded on startup (before collecting active/deployed configs from builder). This might be a breaking change to this repository.
3. Can we break/fork out the repo since we are serving two masters (dialogy template & core slu) at the same time ? every move forward we have to think of backward compatibility.