Add ESS thoreshold logic to SMCFilter

[fritzo]

Addresses #2426

This adds an ESS-based threshold to SMCFilter to decide when to importance resample (resolving a long-standing TODO). The default is to resample when ESS < num_particles / 2. I've also threaded this through the epidemiological CompartmentalModel because I've found a lower threshold can help preserve diversity and find a feasible solution in #2468.

I've also added a couple checks to raise ValueError asap when SMCFilter fails to find a feasible hypothesis.

Tested

• logic updates are covered by existing tests for both SMCFilter and CompartmentalModel
• manually logged ess values to sanity check dynamics
• evaluated improved performance on epidemiological models in Add coalescent likelihoods to contrib.epidemiology #2468

[fritzo]

Derivation:

ess = 1 / sum(p[i] ** 2 for i in range(N))
    = 1 / sum((logp[i] * 2).exp() for i in range(N))
    = logp.mul(2).exp().sum().reciprocal()

[martinjankowiak]

👍