Deprecate MultiHeadAttention

[WindQAQ]

Deprecate MultiHeadAttention since tensorflow/tensorflow@f32c80b is merged.

I'm quite confused what our roadmap is. Should we alias/wrap the functionality to core TF or just remove it? In gelu case, the default argument is changed, and for MultiHeadAttention, the input signature is totally different.

[WindQAQ]

cc @tensorflow/sig-addons-maintainers

[AakashKumarNain]

Yeah there are differences. IMO, we should remove it.

[failure-to-thrive]

Just wondering, is there any policy on backward compatibility?

[bhack]

We have a no merged draft at tensorflow/community#241