Restore the intentional MiscSystemData and generate_trajectory exports

[ChrisRackauckas-Claude]

Ignore this PR until it has been reviewed by @ChrisRackauckas.

What changed and why

MiscSystemData and generate_trajectory were both deliberately exported by the contributors who added them; #4962 reclassified them as accidental internal reexports and removed both exports, which dropped public API in a minor release. This restores the exports, documents both names on the public API pages instead of the internals page, and removes the QA testset that asserted they were internal.

• MiscSystemData — added and exported in b79914d by @AayushSabharwal as the user-facing key for one-off system metadata (setmetadata(sys, MiscSystemData, x)).
• generate_trajectory — added and exported in 5be3f9e alongside the rest of the exported generate_* codegen family (generate_cost, generate_jacobian, generate_control_function, …).

#4979 then filed both docstrings under docs/src/internals/systems.md, whose page banner reads "These APIs may change without notice" — the wrong home for exported API. They move to the pages their siblings live on: MiscSystemData next to the system setmetadata/getmetadata entries in API/System_accessors.md, generate_trajectory into API/codegen.md. The collocation solve method that #4979 filed in the same internals section moves to API/dynamic_opt.md, the page that documents that exact workflow.

ModelingToolkitBase goes 1.67.0 → 1.68.0, in its own commit: widening the public API is a minor bump.

This is the upstream half of SciML/Catalyst.jl#1538, where @AayushSabharwal flagged that MiscSystemData is intentionally exported. No Catalyst follow-up is needed — its LEGACY_DEPENDENCY_REEXPORTS allowlist already carries both names.

Failing before / passing after

MWE run against the working tree, Julia 1.12.4:

using ModelingToolkit
using ModelingToolkit: t_nounits as t, D_nounits as D

for name in (:MiscSystemData, :generate_trajectory)
    println("$name exported from ModelingToolkit: ", Base.isexported(ModelingToolkit, name))
    println("$name public in ModelingToolkit:     ", Base.ispublic(ModelingToolkit, name))
end

@variables x(t)
@named sys = System([D(x) ~ -x], t)
sys = setmetadata(sys, MiscSystemData, "payload")
println("round-tripped metadata: ", getmetadata(sys, MiscSystemData, nothing))

On unmodified master (837ca4d):

$ julia --startup-file=no --project=. mwe.jl
MiscSystemData exported from ModelingToolkit: false
MiscSystemData public in ModelingToolkit:     false
generate_trajectory exported from ModelingToolkit: false
generate_trajectory public in ModelingToolkit:     false
ERROR: LoadError: UndefVarError: `MiscSystemData` not defined in `Main`

With this PR applied:

$ julia --startup-file=no --project=. mwe.jl
MiscSystemData exported from ModelingToolkit: true
MiscSystemData public in ModelingToolkit:     true
generate_trajectory exported from ModelingToolkit: true
generate_trajectory public in ModelingToolkit:     true
round-tripped metadata: payload

Verification

GROUP=QA on this branch, Julia 1.12.4 — the Aqua / ExplicitImports / public_reexports lane, which is what actually asserts that these two names are deliberate API rather than leaked internals:

$ GROUP=QA julia --startup-file=no --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total     Time
QA            |   40     40  4m36.1s
     Testing ModelingToolkit tests passed

Same lane on unmodified master (837ca4d), for the baseline:

$ GROUP=QA julia --startup-file=no --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total     Time
QA            |   52     52  4m26.7s
     Testing ModelingToolkit tests passed

Both are green. The 52 → 40 drop is exactly the deleted testset: 6 assertions (isdefined, !isexported, !ispublic, each against ModelingToolkitBase and ModelingToolkit) × 2 names. Nothing else was removed or loosened — in particular public_reexports still runs and now passes with both names back in REEXPORTED_API.

Docs build (checkdocs = :exports, doctest = true, linkcheck = true) — this is the check that gates the restored exports having rendered docs entries, since an exported name with no @docs entry fails CheckDocument:

$ julia --startup-file=no --project=docs -e 'using Pkg; Pkg.develop([PackageSpec(path="."), PackageSpec(path="lib/ModelingToolkitBase")]); Pkg.instantiate(); include("docs/make.jl")'
[ Info: Doctest: running doctests.
[ Info: ExpandTemplates: expanding markdown templates.
[ Info: CrossReferences: building cross-references.
[ Info: CheckDocument: running document checks.
[ Info: Populate: populating indices.
[ Info: RenderDocument: rendering document.
[ Info: HTMLWriter: rendering HTML pages.
[ Info: Automatic `version="11.39.1"` for inventory from ../Project.toml
docs exit=0

The only warnings are the pre-existing ones on master: size_threshold_warn on three large API pages, the 60 @example text/html notice, and two linkcheck 301 redirects (claytex.com, codecov.io).

Runic (--check --diff, @runic env), typos over the added diff lines, and git diff --check all pass.

Not verified

• Julia LTS (1.10) and prerelease. Everything above ran on 1.12.4. The deleted testset was the only VERSION >= v"1.11"-gated code touched here, so there is no version-specific branch left in the diff, but I did not run the LTS lane.
• GPU, FMI, and downstream/ReleaseTest lanes.
• The other test groups (InterfaceI, InterfaceII, Initialization, …). This diff adds two export lines, moves docs entries, and deletes a QA testset; it changes no solver or codegen behaviour, so I ran the lanes that can actually observe it (QA, docs) rather than the full matrix.

Anything a reviewer should push back on

• Whether generate_trajectory should be exported at all, or only @public. It was exported by its author and every neighbouring generate_* entry point is exported, so this restores it as an export; if the intent was developer-only, @public plus a docs entry would be the alternative and that is a judgement call, not a mechanical one.
• ModelingToolkit's [compat] floor on ModelingToolkitBase is left at "1.64". Only the 1.66/1.67 window lacks these exports, and MTK/MTKBase release in lockstep from this monorepo, so raising the floor to "1.68" seemed like more churn than the two-version gap warrants. Say the word if you would rather pin it.