benchmark_with_validation stream removes attribute 'task_labels'

[wang-xulong]

🐛 Describe the bug
A clear and concise description of what the bug is.
I use benchmark_with_validation stream to get the train_stream and valid_stream. when I use the strategy training , the train_stream or valid_stream don't have attribute 'task_labels'.

"AttributeError: 'DatasetExperience' object has no attribute 'task_labels'"

I check the issue is that the generated streams( train_stream and valid_stream) belong to the EagerCLStream while the test_steam belongs to the NCStream. I infer that the EagerCLStream don't have the attribute 'task_labels'. So the training process cannot be executed.

🐜 To Reproduce

scenario = SplitCIFAR100(
        n_experiences=10,  # 10 incremental experiences
        return_task_id=True,  # add task labels
        seed=1,
        shuffle=True
    )

  scenario = benchmark_with_validation_stream(scenario, validation_size=0.25)
  train_stream = scenario.train_stream
  valid_stream = scenario.valid_stream
  test_stream = scenario.test_stream

strategy = EWC(
        model=model,
        optimizer=optimizer,
        criterion=criterion,
        train_mb_size=64,
        train_epochs=appr_args['epochs_max'],
        eval_mb_size=64,
        device=device,
        evaluator=eval_plugin,
        ewc_lambda=0.4,
        eval_every=1,
        plugins=[lr_scheduler_p]
    )

  # train and test loop
  for train_task in train_stream:
      strategy.train(train_task)                                  << error is here

🐝 Expected behavior
A clear and concise description of what you expected to happen.

    strategy.train(train_task)      can be performed successfully