Bad sharding (all* operations) for non-divisible outputs

Example:

```julia
y_ra = Reactant.to_rarray(
rand(Float32, 8, 8, 4); sharding=Sharding.NamedSharding(mesh, (:x, nothing, :y))
)

function wrap(x)
y = x[:, :, 1:1]
return cat(x, y, dims=3)
end

@code_xla wrap(y_ra)
```

Ideally we detect this and pad the outputs to prevent an all-gather here