Remove Turing integration tests

[penelopeysm]

Closes #703

General breaking changes

• Removed link! and invlink!. They've been deprecated for about two years now: https://github.com/TuringLang/DynamicPPL.jl/blame/f0c31f045ccd9b69374a68ea24d3823d632a5234/src/varinfo.jl#L1323-L1329

Breaking changes in TestUtils.DEMO_MODELS

• What was originally demo_assume_observe_literal has been renamed to demo_assume_multivariate_observe_literal to avoid name clash with new model
• demo_assume_literal_dot_observe renamed to demo_assume_dot_observe_literal to match the naming pattern of the other models
• Added a new demo_assume_observe_literal, which is a univariate assume

@model function demo_assume_observe_literal()
    # univariate `assume` and literal `observe`
    s ~ InverseGamma(2, 3)
    m ~ Normal(0, sqrt(s))
    1.5 ~ Normal(m, sqrt(s))
    2.0 ~ Normal(m, sqrt(s))

    return (; s=s, m=m, x=[1.5, 2.0], logp=getlogp(__varinfo__))
end

This is a slightly annoying change, because anybody who uses demo_assume_observe_literal will have the behaviour changed silently. But the only code that uses it is DynamicPPL and Turing, so I think it's fine, as long as I fix it upstream too. (See GitHub-wide search for demo_assume_observe_literal)

Changes to test-internal utils

• Added a function make_chain_from_prior([rng,] model, n_iters) which constructs an MCMCChains.Chains object by sampling from the prior of modelforn_iters` iterations.

Note this is in test/test_util.jl rather than src/test_utils/foo.jl, hence only accessible inside the test suite.

Changes to test suite

• Removed the old Turing integration tests (see Turing integration tests should be moved to (or included in) Turing repo #703).
• Most of them have been incorporated back into the DynamicPPL test suite itself. The functionality that required Turing.jl was either sampling (which has been replaced with the make_chain_from_prior function above), or some sampling algorithm with some space parameter (which has been replaced with a dummy algorithm MyAlg in test/varinfo.jl).
• The remaining tests that haven't been added back are being tracked here: Update for proposed DynamicPPL 0.32 Turing.jl#2419 However, I think this can be part of a different PR as it would require more extensive modifications to our DemoModels module.

Feedback needed:

• make_chain_from_prior seems too useful to languish in test/test_util.jl. Should this be in DynamicPPLMCMCChainsExt?

[coveralls]

Pull Request Test Coverage Report for Build 12302036826

Details

• 22 of 22 (100.0%) changed or added relevant lines in 1 file are covered.
• 22 unchanged lines in 5 files lost coverage.
• Overall coverage decreased (-0.5%) to 86.025%

---

Files with Coverage Reduction	New Missed Lines	%
ext/DynamicPPLForwardDiffExt.jl	1	66.67%
src/logdensityfunction.jl	1	79.41%
src/threadsafe.jl	2	58.47%
src/varinfo.jl	8	84.72%
src/model_utils.jl	10	19.64%

Totals
Change from base Build 12295360310:	-0.5%
Covered Lines:	3675
Relevant Lines:	4272

---

💛 - Coveralls

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 86.02%. Comparing base (0f07520) to head (ea6fd42).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #733      +/-   ##
==========================================
- Coverage   86.49%   86.02%   -0.47%     
==========================================
  Files          36       36              
  Lines        4272     4272              
==========================================
- Hits         3695     3675      -20     
- Misses        577      597      +20     

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

[penelopeysm]

Some specific notes on the tests that got removed.

I promise that apart from everything I highlighted here, everything did get moved into the DPPL test suite.

[penelopeysm]

The tests in this file (test/turing/compiler.jl) are the only ones that weren't moved somewhere inside the DynamicPPL test suite. That's because they all actually rely on sampling the posterior and performing numerical checks.

Consequently, I've just moved them to Turing, at least for the time being. This is being tracked at TuringLang/Turing.jl#2419

I don't entirely know the purpose of these tests, but I got quite tired of trawling through all of the tests and figuring out what they were doing / whether they were still needed, so I think this is a reasonable compromise for now.

[penelopeysm]

I removed this test without replacing it anywhere else. It seems to be checking that we can sample from a model that has its priors in an array. But I think this aspect is already covered in our DEMO_MODELS, specifically demo_dot_assume_observe_index, so I made the call to remove this.

DynamicPPL.jl/src/test_utils/models.jl

Lines 272 to 282 in 2252a9b

	@model function demo_dot_assume_observe_index(
	x=[1.5, 2.0], ::Type{TV}=Vector{Float64}
	) where {TV}
	# `dot_assume` and `observe` with indexing
	s = TV(undef, length(x))
	s .~ InverseGamma(2, 3)
	m = TV(undef, length(x))
	m .~ Normal.(0, sqrt.(s))
	for i in eachindex(x)
	x[i] ~ Normal(m[i], sqrt(s[i]))
	end

[mhauru]

What do you think is the right order to merge this and TuringLang/Turing.jl#2419?

[penelopeysm]

What do you think is the right order to merge this and TuringLang/Turing.jl#2419?

This one first, as it's not just tests being modified but also source code so the Turing PR will have to handle the breaking changes here.

[mhauru]

This one first, as it's not just tests being modified but also source code so the Turing PR will have to handle the breaking changes here.

Sounds good. Given that this creates a moment when the tests are in neither package, we should make sure to readd them in Turing.jl soon.

Great stuff, thanks @penelopeysm!