DPPL 0.37 compat for particle MCMC

[mhauru]

This may or may not be done functionality-wise, except VariableOrderAccumulator seems to go wrong when using Gibbs. (This is probably the same problem as before in merging accumulators.) Let's see how CI does. The code has some pretty ugly bits though, that should be improved.

FYI @penelopeysm

[github-actions]

Turing.jl documentation for PR #2625 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2625/

[codecov]

Codecov Report

❌ Patch coverage is 6.66667% with 56 lines in your changes missing coverage. Please review.
✅ Project coverage is 24.49%. Comparing base (7ca59ce) to head (6d6fac8).
⚠️ Report is 2 commits behind head on mhauru/dppl-0.37.

Files with missing lines	Patch %	Lines
src/mcmc/particle_mcmc.jl	0.00%	54 Missing ⚠️
src/mcmc/ess.jl	0.00%	2 Missing ⚠️

Additional details and impacted files

@@                 Coverage Diff                  @@
##           mhauru/dppl-0.37    #2625      +/-   ##
====================================================
- Coverage             24.84%   24.49%   -0.35%     
====================================================
  Files                    22       22              
  Lines                  1469     1498      +29     
====================================================
+ Hits                    365      367       +2     
- Misses                 1104     1131      +27     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[penelopeysm]

As discussed earlier it's looking good. Finer details we talked about:

• Having a specific accumulator for the particle weight (i.e., the log-likelihood of the most recent observation). Adding a quantity to this accumulator would be responsible for triggering the resampling (via Libtask.produce).
• This would allow us to accumulate logprior / loglikelihood ( / logjac) as usual, so that they don't have a special meaning inside the pMCMC file (the current code hijacks vi.logp[] for its own purposes).
• Unsure whether @addlogprob! should also trigger resampling. We think it probably should (if it is incrementing the likelihood by a nonzero amount).
• We'll probably leave set_retained_vns_del! there for now.

[penelopeysm]

Turing.jl/src/mcmc/particle_mcmc.jl

Lines 45 to 51 in b0df6a6

	score = consume(trace.model.ctask)
	if score === nothing
	return nothing
	else
	return score + DynamicPPL.getlogjoint(trace.model.f.varinfo)
	end
	end

I realised the call to getlogjoint catches not only @addlogprob! but also Gibbs-conditioned variables (I think).

[mhauru]

Most, though not quite all, of the remaining test failures are because merge and subset don't handle VariableOrderAccumulator. Do you want to merge this first and make that a separate PR?

[penelopeysm]

Quite happy to!

[penelopeysm]

line 337 needs the same fix as per discussion in TuringLang/DynamicPPL.jl#999.

[penelopeysm]

(the fix for this is in #2629)

[penelopeysm]

Just thought to update the comments I left, feel free to apply those changes + merge if you're happy with them.