Add coalescent likelihoods to contrib.epidemiology

[eb8680]

Addresses #2426
Pair-coded with @fritzo

This PR adds a new example model that includes a phylogenetic term.

(part of this PR was moved to #2474:)

This PR adds a phylogenetic likelihood translated from EpiGenMCMC to `contrib.epidemiology`, and adds a new example model that includes a phylogenetic term. There are two implementations of the likelihood here, a vectorized and non-vectorized version (in coalescent event sequence and hence in simulation time). In smoke testing with the example in this PR, the two appear to behave identically in `transition_fwd`, and the vectorized version should be easy (with a bit of judicious masking) to integrate with our parallel-scan inference machinery.

Tasks:

• Add sequential likelihood implementation
• Add naive vectorized likelihood implementation
• Add a temporary script for preprocessing data
• Add a simple example and use it as a smoke test with some real data
• Integrate the vectorized likelihood together with some naively tensorized data into transition_bwd
• Convert likelihoods to TorchDistributions
• Add unit tests that check math and broadcasting
• Add documentation
• Investigate less memory-intensive approaches to vectorized likelihood evaluation, perhaps by converting the data into PyTorch sparse tensors
• Usability: allow non-sorted coal_times or at least explain error
• Usability: consider allowing negative coal_times
• Add unit tests that check against some ground-truth likelihoods (blocked by lack of ground truth)
• Remove temporary preprocessing script with R/Rpy2 dependency

[fritzo]

I'd like to update this PR with math from #2474 and update the example to parse a phylogeny.nwk

[fritzo]

Closing in favor of a tutorial notebook.