arguments, either tensors or resources.

The primary benefit of the pass is to hoist `num_replicas` `_TPUCompile`s
into a single `_TPUCompile`.

This pass assumes that when a `tf.Shape` directly inputs from `replicate`
params, then it is the same shape across replicas.

For example, the following

```mlir
tf_device.replicate([%0, %1] as %ri: tensor<*xi32>) {n = 2 : i32} {
  %2 = "tf.Shape"(%ri) : (tensor<*xi32>) -> tensor<?xi32>