Fix current test errors, update to testthat 3rd edition

[andrjohns]

Several tests were failing due to changes in the outputs from other packages that were being compared against (e.g., lme4) as well as the reliance on deprecated syntax from ggplot2.

I've also updated the actions workflow to use the built-in concurrency handling (instead of the custom n1hility/cancel-previous-runs workflow) and automatically detect and install all suggested deps instead of the hand-maintained list.

This PR also updates the test syntax for testthat 3e, enabling running the tests in parallel.

I've tried to keep the changes grouped by commit so it's easier to review, plus an (AI-generated) summary below, but let me know if it would be better split into multiple PRs!

Summary

This branch gets CI running again and updates the package and its tests for newer versions of its dependencies (lme4, loo, bayesplot, ggplot2, testthat and roxygen2).

• VarCorr.stanreg (R/stanreg-methods.R): each matrix now has class c("vcmat_us", "matrix", "array"), to match what current lme4 returns.
• get_common_cnms (R/jm_data_block.R): the result is now sorted by name, so the grouping-factor order is the same in every submodel of stan_jm/stan_mvmer.
• pp_check (R/pp_check.R): ppc_error_scatter_avg now gets the default x = 1:length(y), like ppc_intervals, ppc_ribbon and ppc_error_binned already did.
• plot_nonlinear (R/stan_gamm4.R):
  • aes_() (deprecated) is replaced by aes(), and size by linewidth for lines and ribbons.
  • Setting the variables to NULL locally removes the R CMD check "no visible binding" NOTEs.
• testthat edition 3, run in parallel:
  • Dropped context().
  • expect_equivalent / check.attributes = FALSE became ignore_attr = TRUE.
  • tol = became tolerance =.
  • expect_is became expect_s3_class, expect_type or inherits().
• loo: the expected error text is now "All models must have the same number of observations". The helper passes r_eff = TRUE to match the implementation.
• posterior_predict: the helper had a typo, regexep, so the error message was never checked. It is now regexp with the current message.
• pp_check: the newer bayesplot functions (ppc_calibration*, ppc_dots, ppc_loo_calibration*, ppc_loo_pit_ecdf) are excluded from the generic loop.
• VarCorr: attribute names are sorted before comparing.
• clogit: the reference values that depended on the StanHeaders version were removed, and ITER goes from 100 to 200.

[andrjohns]

Unfortunately, the mac R-devel workflow will still fail until this PR is merged upstream: r-lib/actions#1108

[jgabry]

Thanks @andrjohns, this is great. I’ll take a closer look soon.