Add flag to drop example from PredicitonResult

RunInference currently returns both the example/input and prediction. 
```

PredictionResult:
  example: torch(...)
inference: torch(...)

```

Users may want the ability to only return the inference to minimize potential memory/serialization issues later in the pipeline. They can do this with a flag, and the return value may look like
```

PredictionResult:
  example: None
  inference: torch(...)

```

Imported from Jira [BEAM-14362](https://issues.apache.org/jira/browse/BEAM-14362). Original Jira may contain additional context.
Reported by: yeandy.
Subtask of issue #21435