Delete nodes that have no uses after declustering.

This may seem pointless because XLA will trivially DCE these, but removing them
can reduce the number of incoming arguments to the XLA computation.  If these
incoming arguments are integer typed then this can often reduce host to device
copies (since inputs to XLA computations live on the device while TF nodes
generally place integer typed tensors on the host).

PiperOrigin-RevId: 220741169