Both the embedding op and compile op are assumed to be wrapped in separate
 tf_device.launch() ops. This is because the embedding op is head outside
 compiled and the compile op is wrapped in launch to execute on host
 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"() ({