Proposal for more generic output head format.

[Tilps]

No description provided.

[Tilps]

Further ideas to consider before moving forward with this.

1. Should NETWORK_SE_AND_EXTENSIBLE_OUTPUT_HEADS imply that filter size and disable_relu are always respected, rather than the comment of different defaults for residual stack vs output heads. I'm leaning towards yes.
2. Should there be an optional 'mapping layer' at the end of the output message. It would be a message containing repeated int32 permutation = 1; Where -1 is to indicate inputs that are discarded and otherwise is the output index to forward to. It sounds good I think... I would assume the microseconds needed to do it in cpu are probably acceptable without needing to optimize too much.