Computing causal quantities across worlds is too clunky

[eb8680]

Can we make it easier to select the correct subsets of batched causal variables to compute some cross-world estimand like a treatment effect? This question seems important enough to be addressed before an initial release.

Currently the API is really clunky and relies on unspecified behavior (the ordering of plate dimensions created by MultiWorldCounterfactual), as can be seen in e.g. the definition of ITE in the SLC example:

Y_cf[..., 1, 0, :, :] - Y_cf[..., 0, 1, :, :]

What kind of higher-level interface could we provide for specifying quantities like this that's less error-prone than indexing into the plate dimensions but more tightly scoped in its implementation than named indexing in Funsor?

One possibility: if we were to add an optional name keyword argument to the intervene primitive, we could define a named indexing operator select_world handled by MultiWorldCounterfactual:

@effectful(type="select_world")
def select_world(value: T, intervention_indices: Mapping[str, int], *, event_dim: int = 0) -> T:
    ...

Then an expression like the one above could be made unambiguous:

a = intervene(a, a_, name="a")
...
b = intervene(b, b_, name="b")
...
select_world(Y_cf, dict(a=0, b=1)) - select_world(Y_cf, dict(a=1, b=0))

Specifying names for intervention sites could be largely automated in practice by modifying the do handler to use the names of pyro.sample sites supplied to it by a user.

[eb8680]

Addressed by #97