Lower computed initialization maps via symbolic tracing

[ChrisRackauckas]

Summary

• Trace ModelingToolkit initialization state/parameter maps with Symbolics variables instead of integer index tokens. Map outputs classify per entry: a bare traced variable stays a static gather recipe (InitializationSourceIndex), a literal becomes an InitializationConstant, and a computed symbolic expression compiles through Symbolics.build_function into a generated function (InitializationScalarExpression / InitializationArrayExpression) evaluated on-device against the same flat source vector. Arrays with any computed entry compile as one static-output function.
• RuntimeGeneratedFunctions.drop_expr strips the stored body::Expr so compiled recipes are isbits and fit inside GPU kernel problems (the same mechanism MTKBase codegen uses).
• Symbolics and RuntimeGeneratedFunctions join the ModelingToolkitBaseExt trigger list — both always load with ModelingToolkitBase, so extension activation is unchanged.
• Raw HomotopyProblem initialization now fails with a clear ArgumentError instead of a MethodError, and the tutorial documents the homotopy(actual, simplified) → actual rewrite (exact by the operator's own numeric semantics: homotopy(a, s) = a) that yields an equivalent supported initialization, with a runnable recipe.
• Tutorial restriction list updated: the "fallback getters are not yet lowered" limitation is lifted; value-dependent branching in maps remains unsupported (untraceable).

The bug this also fixes

The integer-token trace was not merely limited — it was unsound. A computed map output like 2·token + 1 is integer-valued; when it happens to land in 1:source_count it is indistinguishable from a legitimate index, and initialization silently gathers wrong values while reporting success. Reproducer (plain ODE, initialization_eqs = [mx^3 + mx ~ 2, my ~ 2mx + 1], correct init mx = 1, my = 3):

master (silent wrong answer):

COMPAT OK: true
(sol.u[1]).u[1] = [0.0, 1.0]      # wrong, retcode Success

this PR:

COMPAT OK: true
(sol.u[1]).u[1] = [3.0, 1.0]      # correct (state order [my, mx])
compat umap: InitializationStateMap
compat pmap: InitializationParameterMap

Whether master errors or silently mis-initializes depends only on where the aliased value lands: the same equations under different variable naming produce the out-of-range case, which errors. The new Computed initialization maps testset hits that shape on master:

Computed initialization maps: Error During Test ...
  ModelingToolkit initialization maps must copy numeric values from the ODE or initialization problem.
   [2] source_recipe(index::Float64, source_count::Int64)

With this PR both new testsets pass:

Computed initialization maps     |    4      4  35.9s
Homotopy-stripped initialization |    3      3   4.0s

Testing

• GROUP=JLArrays DAE/MTK suite: all 16 testsets pass, including MTK pendulum (19/19) and the SCC testsets from Run MTK SCC initialization sequentially on GPU kernels #511 — the pure-gather fast path is preserved under symbolic tracing.
• QA 27/27, JET 75/75, Testing DiffEqGPU tests passed.
• Runic and typos clean on all changed files.
• Docs: the edited tutorial page builds; the full local docs build fails only on CUDA @example blocks in untouched pages (Pascal code 801 limitation, as on Bump to 3.20.0 for the SCC initialization feature #515) — Documentation CI is the arbiter.
• CUDA not run locally (same pre-existing local Pascal limitation documented in Run MTK SCC initialization sequentially on GPU kernels #511).

Review notes: the map-tracing rewrite touches every recipe path, so the pendulum + Test-3 suites doubling as regression coverage for pure-gather maps is the main safety net. Judgment call worth scrutiny: arrays with any computed entry compile the whole array into one generated function (bare entries become direct indexing) rather than mixing per-element recipes.

---

This PR should be ignored until reviewed by @ChrisRackauckas.

🤖 Generated with Claude Code

Note on Documentation CI: .github/workflows/GPU.yml gates the Documentation job with if: github.event_name == 'push' || !github.event.pull_request.draft, so it shows as skipping while this PR is a draft and only runs once the PR is marked ready for review (or after merge — which is how #511's missing_docs failure escaped to master). Local validation of the edited page is in the Verification section; the full local build cannot go green on this machine (Pascal code 801 on CUDA example pages).