Enzyme: scatter! reverse rule only supports + and -, other ops throw MethodError

`scatter` and `scatter!` accept `+`, `-`, `*`, `/`, `max`, `min` and `mean`, and gradients work for all of them under Zygote. Under Enzyme only `+` and `-` work — the other five throw a `MethodError` in the reverse pass.

The forward pass succeeds, so the failure only shows up when the gradient is taken.

### Reproducer

```julia
using NNlib, Enzyme, Zygote
using Statistics: mean

src = Float32[1.0 0.5 2.0 1.5 0.8
0.7 1.2 0.9 1.1 1.3
1.4 0.6 1.8 0.4 1.0]
idx = [1, 2, 2, 3, 3]

f(op, x, idx) = sum(abs2, NNlib.scatter(op, x, idx))

for op in (+, -, *, /, max, min, mean)
dsrc = zeros(Float32, size(src))
try
Enzyme.autodiff(Reverse, f, Active, Const(op), Duplicated(copy(src), dsrc), Const(idx))
println(rpad(string(op), 5), " ok")
catch e
println(rpad(string(op), 5), " ", first(split(sprint(showerror, e), '\n')))
end
end
```

### Output

```
+ ok
- ok
* MethodError: no method matching reverse(::EnzymeCore.EnzymeRules.RevConfigWidth{1, false, true, (false, false, true, false, false), false, false}, ::Const{typeof(NNlib.scatter!)}, ::Type{DuplicatedNoNeed{Matrix{Float32}}}, ::Nothing, ::Const{typeof(*)}, ::Duplicated{Matrix{Float32}}, ::Duplicated{Matrix{Float32}}, ::Const{Vector{Int64}})
/ MethodError: no method matching reverse(... ::Const{typeof(/)} ...)
max MethodError: no method matching reverse(... ::Const{typeof(max)} ...)
min MethodError: no method matching reverse(... ::Const{typeof(min)} ...)
mean MethodError: no method matching reverse(... ::Const{typeof(mean)} ...)
```

The in-place `scatter!(op, dst, src, idx)` form fails for the same five ops.

Zygote returns correct gradients for all seven ops — I checked them against `FiniteDifferences.central_fdm(5, 1)` and they agree — so this only affects the Enzyme path.

### Where it bites

In GraphNeuralNetworks.jl this blocks Enzyme on every attention layer (they use `scatter` with `max` for the softmax) and on every mean aggregation, which is 8 layers in total.

### Versions

Julia 1.12.5, NNlib 0.9.44, Enzyme 0.13.199, Zygote 0.7.12.

Start with the NNlib.scatter and scatter! entry points and run the provided Enzyme reproducer for all seven operations. Check the reverse-pass handling for the five failing operations, including the in-place form. Done means Enzyme no longer throws for those cases and the resulting gradients agree with the existing Zygote and finite-difference checks.

Written by the indexing model from the issue text.