`withgradient` and `gradient` are incoherent for scalar input

[gdalle]

julia> Tracker.gradient(abs2, 3)
(6.0 (tracked),)

julia> Tracker.withgradient(abs2, 3)
(val = 9, grad = (nothing,))

[mcabbott]

Yes, as documented but a potential surprise:

help?> Tracker.withgradient
  withgradient(f, xs...)

  This computes the value f(xs...) and the gradient with respect to xs. However, it differs from
  gradient in several other respects:

,,,

    •  Only objects satisfying Optimisers.isnumeric are regarded as parameters, thus in
       particular arrays of integers are ignored.

help?> Optimisers.isnumeric
  isnumeric(x) -> Bool

  Returns true on any parameter to be adjusted by Optimisers.jl, namely arrays of non-integer
  numbers. Returns false on all other types.

...

[gdalle]

Okay thanks. The behavior is the same for scalar floats btw, but now I understand why.

julia> Tracker.gradient(abs2, 3.)
(6.0 (tracked),)

julia> Tracker.withgradient(abs2, 3.)
(val = 9.0, grad = (nothing,))

[mcabbott]

Yea I don't know that this is ideal, nor ideally documented. Possibly it's a reason for Flux to own this function.