Bubbles up sharding configuration from `cluster_func` regions into
the attributes of `cluster_func`. This is done by parsing the
`XlaSharding` / `TPUPartitionedOutput` / `TPUPartitionedInput` ops inside
`cluster_func`.

For example, given the following `cluster_func` wrapping `func`:

```mlir
  func @test(%arg0: tensor<*xi32>) {
    "tf_device.cluster_func"(%arg0) {
        func = @func,