during TPURewritePass.

    For example, the tf.OpA with the `mini_batch_splits` attribute will be
    moved after _TPUCompileMlir and the first input will use the
    _TPUCompileMlir program output:

    ```mlir
    "tf_device.launch"() ({
     %cst_0 = "tf.Const"() {value = dense<""> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>
     "tf.OpA"(%cst_0) { mini_batch_splits = ""} : (tensor<1x!tf_type.string>) -> ()

//
// tf_executor.island wraps "tf.OpA"
// tf_executor.island {
//  "tf_device.replicate" {n = 2} {
//    "tf.OpB"
//    "tf_device.parallel_execute"() ({
//      "tf.OpC"
//    }, {
//      "tf.OpD"
//    })
//  }
//
// The above IR will be flattened after `ReplicateToIslandPass` and
// `ParallelExecuteToIslandsPass` as follows:
//
// tf_executor.island wraps "tf.OpA"

    %handle0, %handle0_control = tf_executor.island wraps "tf.SomeOp"() : () -> !tf_res
    %handle1, %handle1_control = tf_executor.island wraps "tf.SomeOp"() : () -> !tf_res
    %control_A = tf_executor.island wraps "tf.OpA"() : () -> ()
    %while_out:2, %while_control = tf_executor.island(%control_A) wraps "tf.While"(
            %handle0, %handle1) {
        body = @body, cond = @cond, is_stateless = false