GAME scoring module improvement

[XianXing]

Currently GAME's scoring module is tightly coupled with evaluation, and the output schema of the computed scores are not user friendly, because the computed scores can't be joined back to the original data for downstream use cases.

We would like to have an improved GAME scoring module with the following features:

• Evaluation will no longer be part of the scoring module, as we shouldn't assume the label/response is available for evaluation purpose for each input example to be scored.
• Input data schema: compared to the Avro schema for training data, the scoring data's schema is almost identical except the following three things:
  1 Each input example is expected to contain a String typed field called uid (for Avro record), which should represent the unique id of each input example to be scored
  2 The response and weight field is no longer required as part of the input
  3 If the offset is included as part of the input, then the final score would be offset + scoreComputedFromModel
• Output score schema: will leverage ScoringResultAvro.avsc found in photon-avro-schemas.

[XianXing]

Related to #47.

[XianXing]

Reopen this issue, because some of our clients at LinkedIn complained the rigidness of the proposed solution.

Will work on a more flexible and easy-to-use scoring module for GAME.