Gradients dropped by `adapt`

Moving y to "gpu" inside loss causes its gradient to be lost:
```julia
julia> using Tracker, JLArrays

julia> JLArrays.allowscalar(false)

julia> Tracker.withgradient((x,y) -> sum(x[1:2] + jl(y))^2, jl([1,2,3.0]), [4,5.0])
(val = 144.0, grad = ([24.0, 24.0, 0.0], [0.0, 0.0]))

julia> ans.grad[1] isa JLArray
true
```
unlike Zygote:
```julia
julia> Zygote.withgradient((x,y) -> sum(x[1:2] + jl(y))^2, jl([1,2,3.0]), [4,5.0])
(val = 144.0, grad = ([24.0, 24.0, 0.0], [24.0, 24.0]))
```

Reproduce the Julia examples using Tracker.withgradient, JLArrays, and Zygote, then trace how the jl(y) value and its gradient are handled during adaptation to the GPU. Done means the Tracker example returns the nonzero gradient for y while preserving the existing x gradient behavior.

Written by the indexing model from the issue text.