Optimize transposed operands / transpose(result) for custom_call

[avik-pal]

module @reactant_traced_... attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<3x2xf32> {enzymexla.memory_effects = ["read", "write", "allocate", "free"]}) -> tensor<2x3xf32> attributes {enzymexla.memory_effects = ["read", "write", "allocate", "free"]} {
    %0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<3x2xf32>) -> tensor<2x3xf32>
    %1 = stablehlo.custom_call @reactant_julia_callback(%0) {api_version = 4 : i32, backend_config = {callback_ptr = 138379625512384 : i64}, has_side_effect = true, operand_layouts = [dense<[0, 1]> : tensor<2xindex>], result_layouts = [dense<[0, 1]> : tensor<2xindex>]} : (tensor<2x3xf32>) -> tensor<3x2xf32>
    %2 = stablehlo.transpose %1, dims = [1, 0] : (tensor<3x2xf32>) -> tensor<2x3xf32>
    return %2 : tensor<2x3xf32>
  }
}

[avik-pal]

With this we can also avoid the confusing manual layout optimizations when lowering (we will need to run this post the lowering passes though)