Clean up exports

[penelopeysm]

As mentioned in #2468, I don't think we need to export so many things. I've done a bunch of cuts here, with the intention that if we don't feel comfortable cutting something, we can easily revert specific commits. In my opinion, though, even though it looks like many sweeping changes, I don't think any of these are hugely controversial on their own.

Concretely, this PR does the following:

1. Stops re-exporting the DynamicPPL module.

2. Stops re-exporting @logprob_str and @prob_str, which were deprecated several months ago (we've killed off deprecated stuff in far shorter timespans)

3. Stops re-exporting DynamicPPL.LogDensityFunction, this is an implementation detail that users probably don't need to construct themselves.

4. Bumps the Distributions.jl compat bound (from 0.23 to 0.25.77) and removes code that was required to support old versions of Distributions. Note that DynamicPPL already requires Distributions 0.25.

5. Removes the blanket re-export of Bijectors and Libtask, as well as the export of Bijectors.ordered. ordered is used in the Gaussian mixture model tutorial, but even there it's imported directly from Bijectors, and I don't think it's unfair to expect users to manually import Bijectors if they want to use this function.

6. Removes the blanket re-export of AbstractMCMC, replacing it with an explicit export list of sample, MCMCThreads, MCMCDistributed, and MCMCSerial. Note that these are the only four identifiers which AbstractMCMC itself exports (https://github.com/TuringLang/AbstractMCMC.jl/blob/master/src/AbstractMCMC.jl) so the only change is that the module name itself (AbstractMCMC) is no longer accessible as Turing.AbstractMCMC.

7. For convenience, re-exports LinearAlgebra.I.

8. Exports DynamicPPL.returned (apparently we never bothered to export it when deprecating generated_quantities, oops haha).

Will let CI run before requesting a review.

[codecov]

Codecov Report

Attention: Patch coverage is 0% with 6 lines in your changes missing coverage. Please review.

Project coverage is 56.43%. Comparing base (24d5556) to head (2409aa1).

Files with missing lines	Patch %	Lines
src/stdlib/distributions.jl	0.00%	4 Missing ⚠️
src/mcmc/Inference.jl	0.00%	1 Missing ⚠️
src/optimisation/Optimisation.jl	0.00%	1 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (24d5556) and HEAD (2409aa1). Click for more details.

HEAD has 15 uploads less than BASE

Flag	BASE (24d5556)	HEAD (2409aa1)
	54	39

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #2474       +/-   ##
===========================================
- Coverage   85.01%   56.43%   -28.59%     
===========================================
  Files          21       21               
  Lines        1582     1561       -21     
===========================================
- Hits         1345      881      -464     
- Misses        237      680      +443     

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

[coveralls]

Pull Request Test Coverage Report for Build 12885951319

Details

• 0 of 6 (0.0%) changed or added relevant lines in 3 files are covered.
• 430 unchanged lines in 11 files lost coverage.
• Overall coverage decreased (-27.7%) to 48.809%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/mcmc/Inference.jl	0	1	0.0%
src/optimisation/Optimisation.jl	0	1	0.0%
src/stdlib/distributions.jl	0	4	0.0%

Files with Coverage Reduction	New Missed Lines	%
src/mcmc/emcee.jl	2	88.46%
src/variational/VariationalInference.jl	4	0.0%
src/mcmc/particle_mcmc.jl	10	80.13%
src/mcmc/Inference.jl	20	75.15%
ext/TuringDynamicHMCExt.jl	29	0.0%
src/mcmc/mh.jl	34	58.96%
src/stdlib/distributions.jl	47	0.0%
ext/TuringOptimExt.jl	50	0.0%
src/mcmc/sghmc.jl	50	20.29%
src/variational/advi.jl	56	0.0%

Totals
Change from base Build 12786420948:	-27.7%
Covered Lines:	758
Relevant Lines:	1553

---

💛 - Coveralls

[mhauru]

I think I agree with all the changes listed in OP, although LogDensityFunction makes me pause and wonder how people might be using it. I don't think I should be the main reviewer here, the more senior members of the team understand user needs better.

Rather than merging straight to master, should we bundle this with #2458 to form a new breaking release to go out when we feel like is the right moment?

[penelopeysm]

Happy to release both PRs together!

I can't say I feel particularly bad about LogDensityFunction tbh, it's still accessible via DynamicPPL with the same functionality, so we're really just asking people to import it from where it should be imported.

[penelopeysm]

There's still the wholesale export of MCMCChains which I don't dare to touch. MCMCChains exports a lot of stuff. https://github.com/TuringLang/MCMCChains.jl/blob/master/src/MCMCChains.jl