conditionally perform convert(concat) -> concat(convert)

```julia
julia> function fn(x)
Float32.(vcat(x[1:4], -x[5:8]))
end
fn (generic function with 1 method)

julia> @code_hlo fn(Reactant.to_rarray(rand(BFloat16, 8)))
module @reactant_fn attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
func.func @main(%arg0: tensor<8xbf16>) -> tensor<8xf32> {
%0 = stablehlo.slice %arg0 [0:4] : (tensor<8xbf16>) -> tensor<4xbf16>
%1 = stablehlo.slice %arg0 [4:8] : (tensor<8xbf16>) -> tensor<4xbf16>
%2 = stablehlo.negate %1 : tensor<4xbf16>
%3 = stablehlo.convert %0 : (tensor<4xbf16>) -> tensor<4xf32>
%4 = stablehlo.convert %2 : (tensor<4xbf16>) -> tensor<4xf32>
%5 = stablehlo.concatenate %3, %4, dim = 0 : (tensor<4xf32>, tensor<4xf32>) -> tensor<8xf32>
return %5 : tensor<8xf32>
}
}
```

This might be good with we are reducing precision but when increasing precision we will end up increasing memory requirements

Reproduce the Julia fn example using Reactant.to_rarray and inspect its @code_hlo output. Start by tracing the handling of stablehlo.convert and stablehlo.concatenate; completion should conditionally use the requested ordering while avoiding increased memory when precision is widened, with coverage for the shown case.

Written by the indexing model from the issue text.