MLFlowServer predict function ignores feature_names parameter

[meoril]

Describe the bug

In MLFlowServer.py, the SeldonComponent's predict function implementation seems to ignore the feature_names parameter, and sends the MLFlow pyfunc predict function a wrong type.
MLFlow's predict function expects DataFrame instance, and the MLFlowServer sends numpy.ndarray instance instead, which causes unexpected behaviours (can't call the columns property of the DataFrame interface).

To reproduce

Serve a simple MLFlow model that needs to access the DataFrame's columns property.
Call the predictions API for your served model with the names and ndarray properties filled with sample data.
Or just look at the MLFlowServer predict function.

Expected behaviour

The registered MLFlow model should receive the X param as a pandas.DataFrame and not as a numpy.ndarray.

Thanks