Add and test a helper for dependent preemptions

[eb8680]

The design in #240 and previous iterations includes a number of new components on which the overall correctness of causation and explanation query implementations depends. To reduce the difficulty of reasoning about these implementations and make incremental progress toward landing causation and explanation queries in master, we should pull out these components and specify and test them in isolation.

One of these components is a helper function for defining dependent interventions that "undo" the effect of upstream chirho.counterfactual.ops.split operations, called undo_split in the draft design in #240

def undo_split(antecedents: Iterable[str] = [], event_dim: int = 0) -> Callable[[T], T]: ...

Dependent interventions generated from undo_split should preserve the indices of their arguments, but the values should behave as if the specified upstream split(s) never happened:

with MultiWorldCounterfactual():
  x_obs: Tensor = ...
  x_cf_1: Tensor = ...
  x_cf_2: Tensor = ...

  x_split = split(x_obs, (x_cf_1,), name="split1")
  x_split = split(x_split, (x_cf_2,), name="split2")

  undo_split2 = undo_split(antecedents=["split2"])
  x_preempted = preempt(x, (undo_split2,), True)

  assert indices_of(x_split) == indices_of(x_preempted)
  assert torch.all(gather(x, IndexSet(split2={0})) == x_preempted)