[core] Support optional secondary outputs

For some operators, there are optional secondary output.
An example is BatchNorm, it has an attribute called training_mode, when this attribute is set to true, there are two extra outputs that appear on BatchNorm: running_mean and running_var.
The model is in the link

I believe this is not yet supported on Aidge.
The problem is that the 2nd and 3rd outputs need to be optional like InputCategory::OptionalData otherwise, there will be a problem when they are not connected in the graph (when training_mode is false).