PadConv2D

[cofri]

This PR introduces a new PadConv2D layer that is the combination of a padding operation and a Conv2D layer. Different padding types can be set:

• "valid": corresponds to the standard "valid" convolution in Keras Conv2D (i.e. no padding). The output shape is smaller than the input shape.
• "same": corresponds to the standard zero padding in Keras Conv2D. The output shape is equal to the input shape.
• "constant": equivalent to "same", i.e. zero padding.
• "symmetric"/"reflect": mirror padding as in tf.pad(). The output shape is equal to the input shape.
• "circular": circular padding. The output shape is equal to the input shape.

Moreover, some bugs are fixed and unit tests are added.

[thib-s]

As PadConv2D is not lipschitz constrained, we need to find a way to make sure that users are aware of it. Should we move it to a separate module for instance ?