Add support for additional intervention types

[eb8680]

In classical causal inference theory, interventions are defined as replacements of constant values with constant values, but many advanced causal queries can be understood as arising from a richer notion of intervention, e.g. replacing one mechanism with another or replacing a value with a distribution.

Fortunately, most of these interventions are fully compatible in principle with the broadcasting-based counterfactual semantics of this library, and queries that make use of richer interventions can therefore be answered with Pyro's existing inference tools.

All that is necessary to implement this support is to extend the definition of the existing intervene primitive and the MultiWorldCounterfactual effect handler to new input and output types.

Intervention types:

• Atomic values: torch.Tensor or numbers.Number
• Dependent interventions: typing.Callable[[T], Intervention[T]]
• Make intervene generic and extensible #113 Interventions on mechanisms: typing.Callable[[T], T] or pyro.nn.PyroModule
• Allow tuple-valued multiple interventions #121 Multiple simultaneous interventions: typing.Tuple[T, ...]
• Stochastic interventions: pyro.distributions.TorchDistribution or pyro.nn.PyroSample
• Free parameter interventions: pyro.nn.PyroParam