General support of files path in dataframe

[yinweisu]

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