LowerExtend with `lhs > size(x, dim)`

```
%8 = "enzymexla.extend"(%6) <{dimension = 0 : i64, lhs = 2 : i64, rhs = 1 : i64}> : (tensor<1x64x1xf32>) -> tensor<4x64x1xf32>
```

```julia
using NeuralOperators, Reactant, Lux, Random, Test

const xdev = reactant_device(; force=true)

model = GridEmbedding([(0.0f0, 1.0f0) for _ in 1:2])
ps, st = Lux.setup(Random.default_rng(), model) |> xdev;

x = rand(Float32, 5, 5, 3, 4) |> xdev

@code_hlo optimize = false model(x, ps, st)
```

Start by running the Julia reproduction with NeuralOperators, Reactant, Lux, Random, and Test, then inspect the generated HLO for the enzymexla.extend operation. Compare the lhs value with size(x, dim); done means the lowering handles lhs greater than the input dimension without producing an invalid extension.

Written by the indexing model from the issue text.