Add a L2 norm distance layer.

[lcy-seso]

The similarity metric is very useful. Currently, we only have cosine similarity, which is not enough. The l2 norm distance is required by one of our users.

• Due to lacking an inner product layer, the l2 distance cannot be constructed by combining some primary math layers.
• Implementing the L2 distance in one layer is more time and memory efficient than the combination of some primary math layers.