Support conversion of Transforms and TransformedDistributions to and from Funsors

[eb8680]

Addresses #309, #386. Blocked by #387, #388, #389.

This PR adds support for converting Transforms and TransformedDistributions to and from Funsors, and for sampling and scoring withfunsor.Distributions with transforms in their value slot.

The latter is achieved with a new term I've called Lebesgue for want of a better idea, which computes log-det-jacobians of lazy expressions substituted into its free variable using funsor.delta.solve. (Update: I'm moving Lebesgue to another PR, since it no longer seems necessary for the basic task of converting TransformedDistributions.)

There are several ways this functionality could have been instantiated. The design choices in this version were made with a view toward locality of changes - the overall distribution API is left unchanged, as is the behavior of normalize and eager, and I've reused funsor.delta.solve even though it should probably be broken up into separate inversion and linearization transformations.

The number of transforms currently supported is very limited, partly because we don't have many transforms wrapped as TransformOps in Funsor and partly because solve cannot handle inverting more complex expressions. I've at least tried to get the two most important higher-order transforms ComposeTransform and InverseTransform working, so that adding more transforms later will be relatively straightforward. For ease of review, I have also chosen to leave support for the JAX backend to a followup PR, though it should simply involve copying the additions to funsor.torch.distributions verbatim.

Tested:

• Added several TransformedDistribution test cases using the infrastructure from Automate distribution testing #389

Tasks remaining:

• Add unit tests for density
• Add unit tests for cons-hashing
• Add unit tests for sampling
• Add unit tests for transform conversion
• Get generic conversion tests to pass

Out of scope for this PR:

• Support more transforms
• Add similar capability to the JAX backend
• Add more distribution conversion unit tests, especially interactions with Independent - added one Independent case, but conversion failing
• Add better documentation for working with transformed distributions and Lebesgue - moving Lebesgue to another PR

[fritzo]

This looks weird. Can you explain what's going on? When is value_name != "value"? The assertion in self.__init__() suggests value_name == "value" IIUC.

[eb8680]

value_name can potentially be anything, since self.value is a lazy expression. This logic is meant to solve the problem of identifying the value name when self.value is not a Variable or even has more than one input (which happens when constructing the Funsor version of a TransformedDistribution).

The simplest nontrivial example of the latter case would be an affine or power transform where the parameters are funsor.Tensors with nontrivial .inputs, although these are not handled in this PR since funsor.delta.solve does not yet support inverting such expressions.

The main use for value_name in this PR is in Distribution.unscaled_sample, which needs to know value_name to construct a sample Delta with the correct .inputs.

[fritzo]

Thanks for addressing nits!