General support of files path in dataframe

Currently only image files are handled and tested with bytes array encode. This is useful for inference endpoint though not too practical for batch inference. We should consider a more general support, i.e. upload all files to s3, substitute path, and reverse the process in inference container