Implement a Joint funsor as a normal form / state machine

[fritzo]

~~Blocked by #51 ~~

It appears that the simplest way to implement an inference strategy is via a Joint funsor that keeps track of various Tensor, Gaussian, and Delta factors in a canonical, searchable data structure. This Joint object is really a state machine / interpreter disguised as a Funsor. Its denotational semantics are defined by its .eager_subs() method, and its operational semantics are defined by the transition operations we call on Joint objects during inference:

• += a log prob term with either a lazy or eager value
• .reduce(ops.logaddexp, -) out variables when they are no longer live
• .reduce(ops.sample, -) variables when they need a ground value (e.g. for control flow)
• .reduce(ops.add, -) out plates when they are complete (see log_joint in minipyro.py)

Note that this can be sensibly implemented even before we implement sampling #52

Tasks

• Implement a Joint funsor with Tensor, Gaussian, and various Delta terms
• Implement += Tensor transition
• Implement += Gaussian transition
• Implement += Delta transition
• Implement .reduce(ops.logaddexp, -) to eliminate random variables
• Implement .reduce(ops.sample, -) to sample random variables
• Implement .reduce(ops.add, -) to eliminate plates variables
• Implement += Joint to fuse two state machines
• Add tests
• Use Joint in examples