Optimizing scalars

It seems like Optimisers only works with vectors. 
```
struct WithFloat
    val :: Float64
end
@functor WithFloat
```
The call to `Optimisers.trainable(WithFloat(4))` returns `(val = 4.3,)`, but `destructure(WithFloat(4))` produces an empty vector.

Would you accept a pull request allowing `destructure` to handle scalar parameters? 