Expose onnx optimizer

[SleepProgger]

This PR exposes the onnx.optimzer and the do_constant_folding parameter of torch.onnx.export to be able to optimize the model after exporting.

The use_optimizer parameter of pytorch_to_keras can be one of:

• True to run all optimizer
• False to run NO optimizer
• An iterable acting as whitelist containing the optimizer names to run (Use onnx.optimizer.get_available_passes() to list all available)

There are only 3 optimizer i had problems with:

• extract_constant_to_initializer which doesn't seem to work well together with onnx2keras, but that should be fixable.
• split_init which lead to an output being a numpy array. Might be fixable in onnx2keras
• split_predict which completly destroyed the model and i am not really sure why to be honest.

Both do_constant_folding and use_optimizer default to False.

[gmalivenko]

Yes, I think it's a good idea to optimize graph before calling the converter.