In FunctionalizeCond, cluster Merge nodes by (CondId, AncestorId, predicate).

Original implementation cluster Merge nodes by (CondId, AncestorId), which is not enough. For the following case:
  pred1 = array_ops.placeholder(dtypes.bool, name='pred1')
  pred2 = array_ops.placeholder(dtypes.bool, name='pred2')
  cond1 = control_flow_ops.cond(pred1, ...)
  cond2 = control_flow_ops.cond(pred2, ...)
  cond3 = control_flow_ops.cond(pred1, use cond1 and cond2)
  cond4 = control_flow_ops.cond(pred2, use cond1 and cond2)

cond3 and cond4 have the same CondId (empty) and the same AncestorId ({pred1, PRED}, {merge_of_cond1, MERGE}), but they should not be merged into one "If" node (because they have different predicates).

This does not happen before cl/231844072, because back then AncestorId uses Switch node in AncestorId, and cond3/cond4 has different Switch nodes.

PiperOrigin-RevId: 232931242