cuda4dnn(concat): write outputs from previous layers directly into concat's output

* eliminate concat by directly writing to its output buffer

* fix concat fusion not happening sometimes

* use a whitelist instead of a blacklist