Whats the reason for the derivative / gradient difference? 

Is there a reason to have a separate function `derivative` for the `gradient` wrt a scalar value? I find that it tends to make me have to add special cases where code would just work for both vectors and scalars if there was a definition of something like:

```julia
AD.gradient(ad::AD.AbstractBackend, f, x::Number) = AD.derivative(ad, f, x)
```