Populate last_hidden_state in the output of feature extraction model class

[zhenchaoni]

Fixes a regression introduced from #805

In #805, we removed the last_hidden_state output from the model class output and return whatever output provided by the underlying onnx model.

But this causes a regression. The composite model class encoder_decoder expects the encoder model to have last_hidden_state output. This output is treated as the encoded values. However, some models doesn't have this output name. Therefore, the encoder decoder loop will fail.

The fix is to provide last_hidden_state if the underlying onnx model uses a different output name.