Implement one-step correction for automated doubly robust estimation

[agrawalraj]

Blocked by #371

Given efficient_influence_function we can implement one step correction as follows (up to some potential interface changes to efficient_influence_fn):

def one_step_correction(
  model: Callable[P, T],
  guide: Guide[P, T],
  target: Functional[P, T],
  test_data: Dict[str, Observation[T]],
) -> torch.tensor:

	eif_fn = efficient_influence_fn(model, guide, target)
	return eif_fn(test_data).mean(axis=0)

[agrawalraj]

Closed by #467