Add default (identity) activations to the Dense layer initializers.

[saeta]

This change follows along with the defaulted values for Conv and other layers.

[rxwei]

I was intentionally not doing so because I wanted the initializer with the activation parameter to be the first candidate in code completion. But for consistency with Keras, I think this change is fine.

[dan-zheng]

LGTM. Perhaps identity activation would be better represented as an optional (no activation) in the future.

(Side note: related issue about identity activation and debuggability of gradient tensor seed shape mismatch.)

[rxwei]

(Side note: related issue about identity activation and debuggability of gradient tensor seed shape mismatch.)

You can check these pretty easily by inserting a precondition or assert in each primitive VJP's pullback to make sure the seed matches the shape of the output. However, that would cause the pullback to capture the original result, which is bad for memory usage. The proper way is to make the differentiation transform insert these checks based on the optimization level.