avoid calling `handle_output` twice per Dagstermill output

[sryza]

If someone yields an Output from within their Dagstermill notebook, the following happens in sequence:

• Within the notebook process:
  • We invoke handle_output to write it to persistent storage
• Within the Dagster step process that launched the notebook process:
  • We invoke load_input to read it from persistent storage
  • We invoke handle_output to write it to persistent storage

The reason we do this is that we need to bubble up an Output to the step worker framework code to let it know that an output was yielded, so it can write out the appropriate events to the instance.

One way to fix this would be to offer a way for the dagstermill code to say "trust me, I'm handling the output" and then have the framework code write out the appropriate events without directly touching the output object.

[rahuldave]

This sounds like a great idea!