Avoid splitting up varnames until absolutely necessary

[penelopeysm]

MCMCChains.jl doesn't like vector-valued or matrix-valued variables. Consequently, in Turing we use DynamicPPL.varname_and_value_leaves to split things like x up into x[1], x[2], ...

This is in general a lossy operation since x cannot be reconstructed from its individual elements.

This conversion used to be done when constructing a Transition. However, there was no need to do it so early. This PR changes it such that it is only done within the _params_to_array method, which is called inside bundle_samples.

On its own, this refactor is quite meaningless from the Turing point of view. However, what it does mean is that if we create a new chains type that doesn't need to break variables up, we can define an AbstractMCMC.bundle_samples(::Vector{<:Transition}, ..., ::NewChainType) which retains the original data structure in the NewChainType, because now the Transition will have that richer information. This has implications for the rewrite of MCMCChains that I'm currently working on.

[github-actions]

Turing.jl documentation for PR #2632 is available at:
https://TuringLang.github.io/Turing.jl/previews/PR2632/

[coveralls]

Pull Request Test Coverage Report for Build 16550602808

Details

• 18 of 18 (100.0%) changed or added relevant lines in 3 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.09%) to 84.997%

---

Totals
Change from base Build 16444793035:	0.09%
Covered Lines:	1252
Relevant Lines:	1473

---

💛 - Coveralls

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.88%. Comparing base (23b92eb) to head (61de0a7).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2632      +/-   ##
==========================================
+ Coverage   84.78%   84.88%   +0.09%     
==========================================
  Files          22       22              
  Lines        1466     1475       +9     
==========================================
+ Hits         1243     1252       +9     
  Misses        223      223              

☔ 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.

[sunxd3]

No objection in principle.

Failed tests look unrelated, but may still worth a look?

[penelopeysm]

Yupppp, failures are due to Libtask not working on 1.12 yet. Thanks!