fix: #991

This issue is actually not because of wrong computation but the absence of activation when loading model. And this is because the previous design separates Activation and ActivationName in LayerArgs. Therefore when loading the model, the ActivationName is assigned but Activation not.

This PR made the following changes:

1. Use a class to wrap the activation function and provide a method Apply to make it similar to calling a layer.
2. Add customized json converter for Activation to make it work with model loading.
3. Benefited from 1, tf.keras.activations can be added to align with tensorflow python.