Native particle MCMC (SMC/PG), revised: dropping AdvancedPS

[yebai]

SMC and PG/CSMC now run directly on Libtask and DynamicPPL. Resamplers are
types, stratified by default; conditional sweeps draw ancestors multinomially.
Philox seeding makes draws reproducible across Julia versions. Chains carry
log_normalizing_constant. Roughly twice as quick as main, allocating half.

Builds on #2848. Supersedes #2853.

Fix #2781
Fix TuringLang/AdvancedPS.jl#110
Fix TuringLang/AdvancedPS.jl#39
Fix TuringLang/AdvancedPS.jl#6
Close #2848. Close #2853.

[github-actions]

Turing.jl documentation for PR #2859 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2859/

[codecov]

Codecov Report

❌ Patch coverage is 94.26230% with 14 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.90%. Comparing base (c55358d) to head (7e3f1ff).

Files with missing lines	Patch %	Lines
src/mcmc/particle_mcmc.jl	94.04%	14 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2859      +/-   ##
==========================================
+ Coverage   85.10%   85.90%   +0.79%     
==========================================
  Files          23       23              
  Lines        1517     1603      +86     
==========================================
+ Hits         1291     1377      +86     
  Misses        226      226              

☔ View full report in Codecov by Harness.
📢 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.

[sunxd3]

The core logic looks good to me.

Agent found two minor issues I think worth a second look. They are unlikely to impact user too much though.
Here are the MWEs.

using Turing, Distributions, StableRNGs, Test

@model function constrained_posterior()
    x ~ Normal()
    @addlogprob! (; logprior = x >= 0 ? 0.0 : -Inf)
    0.0 ~ Normal(x, 1)
end

smc = sample(
    StableRNG(1), constrained_posterior(), SMC(), 1_000; progress=false
)
pg = sample(
    StableRNG(1), constrained_posterior(), PG(10), 200; progress=false
)

for chain in (smc, pg)
    xs = vec(collect(chain[:x]))
    logjoints = vec(collect(chain[:logjoint]))

    @show count(x -> x < 0, xs)
    @test all(x -> x >= 0, xs)
    @test all(i -> xs[i] >= 0 || logjoints[i] != -Inf, eachindex(xs))
end

SMC: 507 / 1000 negative draws
PG:   95 / 200 negative draws

logprior didn't affect particle weight (loglikelihood is fine otherwise)

---

using Turing, Distributions, MCMCChains, StableRNGs, Test

@model function one_observation()
    x ~ Normal()
    0.0 ~ Normal(x, 1)
end

chain = sample(
    StableRNG(1),
    one_observation(),
    SMC(),
    32;
    chain_type=MCMCChains.Chains,
    progress=false,
)

ess = vec(collect(chain[:ess_per_step]))

@show names(chain, :internals)
@show ess
@test any(x -> !ismissing(x), ess)

names(chain, :internals) =
[:log_normalizing_constant, :ess_per_step,
 :logprior, :loglikelihood, :logjoint]

ess =
Union{Missing, Float64}[missing, missing, ..., missing]

agent explained that this is because ess_per_step being vector will be replaced by missing

[yebai]

Thanks, @sunxd3 -- both fixed.