Make DepthwiseConv2D non-clusterable

[arovir01]

This PR removes tf.keras.layers.DepthwiseConv2D from the ClusteringRegistry, as our experimental results suggest that depthwise convolution layers are not suited for clustering.

[arovir01]

@Ruomei @akarmi Please review.

[Ruomei]

Since this is already unit tested, please let me know whether it works for mobilenet tests. Also, in the model-optimization repo, try "grep" DepthwiseConv2D or lower cases to see whether there are any other places that need to be cleaned up. For instance, in RFC, there is this "except DepthwiseConv2D layers"? do we want to keep it?

[akarmi]

Since this is already unit tested, please let me know whether it works for mobilenet tests. Also, in the model-optimization repo, try "grep" DepthwiseConv2D or lower cases to see whether there are any other places that need to be cleaned up. For instance, in RFC, there is this "except DepthwiseConv2D layers"? do we want to keep it?

I would argue that we should keep the mentioning of DepthwiseConv2D in the RFC, as this makes it clear that not all layers with potentially clusterable parameters (weights) are actually clustered. Even more, this PR arguably makes the description of clustering configuraiton in that table in RFC more accurate, as it enables appying cluster_weights() to the whole model, and not per-layer, to reproduce the results in those particular rows.

[akarmi]

@alanchiao, please have a quick look. We decided to disable clustering DepthwiseConv2D layers by default as we have not seen any cases when it works well so far.

[alanchiao]

Seems fine to me. Once weight clustering is launched, these types of changes will be harder to make when considering backward compatibility.

While I wasn't there for the specific CL discussion (and don't know if how intentional it was), pruning also disables DepthwiseConv2D by default.