Fix workflow tag concurrency checks

[ChrisRackauckas-Claude]

Ignore this PR until reviewed by @ChrisRackauckas.

Summary

• replace invalid github.refs references in Downstream, Documentation, and ReleaseTest workflows
• use github.ref_type to keep master and tag runs exempt from cancellation as documented

Root cause

Clean-master reproduction with actionlint reports that github.refs is undefined. A focused git bisect identified 80f44bc as the first commit introducing the Downstream typo. A literal github.ref comparison against refs/tags/* would not implement wildcard matching, so the corrected condition checks ref_type instead.

Validation

• actionlint -color .github/workflows/Downstream.yml
• actionlint -color .github/workflows/Downstream.yml .github/workflows/Documentation.yml .github/workflows/ReleaseTest.yml
• Julia Runic --check over the repository

All commands exited successfully locally on Julia 1.12.6.

[ChrisRackauckas-Claude]

Investigation scratchpad:

1. Reproduced on clean ef2225b master: actionlint exits 1 at Downstream.yml:15 because github.refs is undefined.
2. Ran a focused bisect whose predicate matched only the github.refs diagnostic; first bad commit is 80f44bc.
3. Found the identical condition in Documentation.yml and ReleaseTest.yml.
4. Corrected all three to use github.ref_type, since comparing github.ref literally to refs/tags/* would not perform glob matching.
5. Verified the exact reported command, all three touched workflows together, and repository-wide Runic check; each exits 0.

[ChrisRackauckas-Claude]

CI classification: Catalyst/Modeling failed with the same clean-master downstream regression already present in default run 29390528140: UndefKeywordError parent_sys at Catalyst symbolic_stoichiometry.jl:235 (20 passed, 1 errored). The focused fix is SciML/Catalyst.jl#1509. This workflow-expression-only PR does not touch Catalyst or ModelingToolkit runtime code, so I am not mixing that separate fix here.

[ChrisRackauckas-Claude]

CI monitor update: tests / QA (julia 1, ubuntu-latest) did not report a test failure. Its check annotation says the self-hosted runner lost communication with GitHub while the test step was active (job 87293659149). This is an infrastructure failure, separate from this workflow-expression fix. The two Catalyst Modeling failures remain the known default-branch downstream failure already tracked by Catalyst.jl #1509.

[ChrisRackauckas-Claude]

CI monitor update: ModelingToolkitStandardLibrary.jl/Core/1/ has now failed with the same default-branch AccelerationSensor / mass₊v(t) structural-singularity error tracked by ModelingToolkitStandardLibrary.jl #489. I also reproduced that exact failure locally on clean MTSL master against current MTK master (GROUP=Core: 15 pass, 1 error in translational.jl; overall exit 1). Current failure classes remain two known downstream regressions plus the separate QA runner-loss annotation; no failure points to these expression edits.

[ChrisRackauckas-Claude]

Terminal CI update: 58 checks passed and 4 were skipped. The five failures are all unrelated to this workflow-expression-only change:

• Two duplicate Catalyst Modeling failures are the known current-base failure covered by Test symbolic jump affects through public API Catalyst.jl#1509.
• ModelingToolkitStandardLibrary Core is the known current-base failure covered by Initialize translational acceleration sensor test ModelingToolkitStandardLibrary.jl#489.
• QA lost its self-hosted runner connection: https://github.com/SciML/ModelingToolkit.jl/actions/runs/29396991843/job/87293659149
• ModelingToolkitBase downgrade also lost its self-hosted runner connection after its test step ran for about two hours: https://github.com/SciML/ModelingToolkit.jl/actions/runs/29396991783/job/87293166606

Both disconnects are tracked under SciML/.github#116. The relevant local validation remains clean: targeted/repository-wide actionlint, repository-wide Runic, and diff checks all exited 0.