`ForwardDiff.gradient` is slow

[yebai]

It seems the new gradient function based on ForwardDiff.gradient is very slow, e.g. on the model Turing/benchmarks/naive.bayes.run.jl.

[ChrisRackauckas]

Should probably use in-place gradient functions. But reverse mode is probably more efficient here, so you should probably follow the developments of Cassette.jl

[yebai]

Closed by #485