scalar -> vector and vector -> scalar control flow

Currently derivatives of functions with scalar inputs and vector outputs are called gradients, and the control flow is mixed in with
vector -> scalar functions.
In my opinion, this is confusing and error-prone.

Additionally, for the former an in-place `f!` can make sense while for the latter it does not.