Official support for onnx format.

There's an exporter for XGBoost in https://github.com/onnx/onnxmltools . As far as I'm aware of there's no integration tests on XGBoost's side nor any formal document in XGBoost for this ONNX exporter. Last time I look into it was for fixing an issue in 1.0.0 release candidate, which broke the onnx exporter. But right now the converter should already be in good shape for most of the use cases.

My proposal for future work on XGBoost's side is we support the converter in onnxtools by:
- Change their parser implementation into JSON model format based. As the JSON will be our future representation for many stuffs including internal syncing/check pointing, also it's the base of our parameter validation and the serialization format of our configuration:

``` python
booster.save_config() # returns a JSON document showing only our internal configuration.
```
These use cases can not be replaced by protobuf.

- Add tests. As always this is the best way to reduce chances of accidental breaking changes. There are some corner cases where we should look into during testings. For example, is the global bias correctly converted for logistic models and exponential models? Is there any floating point precision loss during the transformation?

- Mention it in our documents.

- (Optional), integrate it in our `save_model` function to make the converter bidirectional.

@sandys @kylejn27 Let's gather the onnx related discussion here. ;-)