Doesn't support non-tensor output with sklearn-onnx in onnxruntime server

Nobody has claimed this yet.

**Describe the bug**
A sklearn randomforest model converted to onnx with convert_sklearn() throws following error with API based inference:
`[error] MLValueToTensorProto() failed. Output name: output_probability. Error Message: Don't support Non-Tensor values`
While it works fine with python onnxruntime.InferenceSession()

**Urgency**
There is a use-cased being blocked due to this issue. Timelines for model deployment in onnxruntime is suppose to be EOBW

**System information**
- OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
- ONNX Runtime installed from (source or binary): binary
- ONNX Runtime version: 1.0.0
- Python version: 3.7
- Visual Studio version (if applicable):
- GCC/Compiler version (if compiling from source):
- CUDA/cuDNN version:
- GPU model and memory:

**To Reproduce**
Follow the steps till creating onnx model from sklearn randomforest (iris dataset) given in official documentation: http://onnx.ai/sklearn-onnx/
Use the onnx model in onnxruntime for inference.

**Expected behavior**
The output_probability from sklearn model converted to onnx should be a tensor by default

**Screenshots**

**Additional context**
Add any other context about the problem here. If the issue is about a particular model, please share the model details as well to facilitate debugging.

Reproduce the sklearn random forest conversion from the sklearn-onnx documentation with ONNX Runtime 1.0.0, comparing Python InferenceSession with API-based inference. Start at the MLValueToTensorProto failure for output_probability; done means the converted model's output is accepted by API inference as a tensor.

Written by the indexing model from the issue text.