Eliminate need for CompartmentalModel.transition_bwd()

[fritzo]

Addresses #2426

This follows ideas from #2513 to replace the .transition_bwd() method with a conditioned call of the .transition_fwd() method, significantly reducing user effort to create a new model. The additional bit of logic to reverse the flow computation is implemented in a new .compute_flows() method that by default assumes a linear flow terminating in "R", but can be overridden by users to perform arbitrary computations. Only one such override is needed by our current examples.

The new .transition() method is now called under more interpretations, so various logic needs to be generalized:

• t < self.duration now also needs to check for slice instances
• We now need to use ExtendedBinomial everywhere, not only in .transition_bwd()
• The SMC wrapper now needs to provide approximate variables like I_approx.

Tested

• refactoring covered by existing tests
• manually ran a biggish example and verified sane posterior

[fritzo]

This is the one model where we need a custom .compute_flows() method.

[martinjankowiak]

great simplification!