Support nested layer customization and enable better 'in_channels' exception raise

[ChrisWu1997]

Checklist

• I've tested that my changes are compatible with the latest version of Tensorflow.
• I've read the Contribution Guidelines
• I've updated the documentation if necessary.

Motivation and Context

1. Support nested layer customization.
2. Enable better 'in_channels' exception raise in dynamic model definition.

Description

I briefly list the design below just for a reminder.

• Nested layer customization
  • Logic
    Each layer maintain a private attribute self._layers to store possible nested layer. When a Layer is registered as an attribute of another Layer, it is automatically added to the second layer's self._layers. Weights are collected recursively.
    Inner layer set to be unaware of LayerNode concept because it does not need to be involved in the model node graph. Inner layer is only regarded as the outer layer‘s 'private' attribute, which is not known by the model.
  • Example

            class MyLayer(tl.layers.Layer):
    
            def __init__(self, name=None):
                super(MyLayer, self).__init__(name=name)
                self.input_layer = tl.layers.Dense(n_units=20)  # it's ok not to pass `in_channels` here
                self.build(None)
                self._built = True
    
            def build(self, inputs_shape=None):
                self.W = self._get_weights('weights', shape=(20, 10))
    
            def forward(self, inputs):
                inputs = self.input_layer(inputs)
                output = tf.matmul(inputs, self.W)
                return output

    MyLayer would contain three weights: W of MyLayer, W and b of the inner Dense.
• 'in_channels' exception in dynamic model
  Now it is checked when setting Layer as an attribute of dynamic Model. If in_channels not provided, there will be an exception raised, e.g.

  The registered layer `dense_4` should be built in advance. Do you forget to pass the keyword argument 'in_channels'?