Provide a generic mlflow pyfunc wrapper

When using [mlflow](https://mlflow.org/) to train and register a text classifier model using the package, we often have to write our own pyfunc wrapper, which could be tedious.

The idea would be to provide a generic wrapper (with some hypothesis on the stored artifact) usable by all package users.

As a bonus, it could be really great if this wrapper was self sufficient (no need to install the package for inference, only torch and numpy)