feature: add selectable inference content for csv, json, jsonlines, and recordio-protobuf

[wiltonwu]

Description of changes:

• Adds new selectable inference content feature for accept types: application/json, application/jsonlines, application/x-recordio-protobuf, and text/csv
• This new selectable inference content is turned on via the SAGEMAKER_INFERENCE_OUTPUT enviornment variable.
  • If that env var is present, the env var will determine the content in the inference response in a string of content keys separated by commas corresponds to the allowed selectable content
  • e.g. SAGEMAKER_INFERENCE_OUTPUT='predicted_label,probabilities'

Tested with tox and integration tests.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[ericangelokim]

It is alarming that we are now dictating what the correct values to return to customers are, for example get label for binary_log. The decisions you made make sense to me, but we might want to involve product to make sure this behavior is ok to push out.

Also, were integration tests run with the built image?

[wiltonwu]

It is alarming that we are now dictating what the correct values to return to customers are, for example get label for binary_log. The decisions you made make sense to me, but we might want to involve product to make sure this behavior is ok to push out.

Okay! Let's sync up on the product point. Most of the "correct value" decisions are just an extension of the xgboost sklearn API

Also, were integration tests run with the built image?

Yep! Correct

[ericangelokim]

Two small nit comments, LGTM