feat: accept array equations on the NonlinearProblem path

[utkuyilmaz1903]

Summary

• Accept unscalarized array equations in NonlinearProblem and NonlinearLeastSquaresProblem after complete when unknowns are scalar.
• Assemble mixed scalar and array equations into one residual, with one row per array element.
• Preserve out-of-place ForwardDiff.Dual values from unknowns and tunable parameters.
• Expand array derivatives before steady-state conversion.
• Preserve FunctionWrapperSpecialize and non-square residual prototypes.

Constraints

• Array unknowns remain unsupported.
• Array residuals require mtkcompile before using jac = true or sparse = true.
• The out-of-place allocator is restricted to nonlinear codegen. DAEProblem{false} Dual promotion is outside this PR.

Tests

• 1D and 2D array equations
• Mixed scalar and array equations
• In-place, out-of-place, expression, and specialization paths
• Non-square least-squares residuals
• Unknown and parameter AD
• Constant-size generated code up to n = 101
• Existing feat: accept array equations on the implicit-DAE path #4983 DAE regression suite

Compat

• New tests import unwrap from Symbolics so they run on older Symbolics.
• DiffEqBase floor is 7.18.1 so downgrade CI accepts already-unwrapped parameters in nested AutoDespecialize.

Related: SciML/MethodOfLines.jl#623.