Add support for transformers with arbitrary request/response format (ie. not SeldonMessage)

[mlegore]

One feature I have been looking for in Seldon is the ability to pre-process requests of formats that don't comply with SeldonMessage. This would be useful for glueing together systems that don't know they're calling seldon.

For example, I'd like to set up a model to be called from a Google Pub/Sub message, but the json message that is sent by Pub/Sub is not in the format of a SeldonMessage, so right now I would have to set up another worker component (a cloud function or similar) to translate the message into a SeldonMessage before calling Seldon. It would be useful if seldon could include that in its routing graph and keep more of the logic contained within Seldon.

[axsaucedo]

@mlegore although not exlplicitly documented it is currently possible to send arbitrary payloads to the API, and process them using predict_raw, or in your case transform_input_raw. We currently have #2497 to explore this, so closing this issue.

[mlegore]

@axsaucedo is there some examples or documentation for this, in particular how to do this with a language other than python?