On Surjectivity of Neural Networks: Can you elicit any behavior from your model?