fix: don't leak the operating point into unbounded nonlinear bounds

[ChrisRackauckas-Claude]

Please ignore this PR until it has been reviewed by @ChrisRackauckas.

What changed and why

generate_nonlinear_bounds resolves symbolic bounds (e.g. bounds = (some_parameter, 1.0)) by merging the operating point into the bounds map. The map is seeded only for finite bounds:

for (var, b) in zip(dvs, lb)
    b === -Inf && continue                      # unbounded unknowns get no entry
    write_possibly_indexed_array!(lbmap, var, Symbolics.SConst(b), COMMON_NOTHING)
end
left_merge!(lbmap, op)                          # ...so `op` supplies them — as u0 values
lb = varmap_to_vars(lbmap, dvs; tofloat = false)

so the merge also hands every unbounded unknown its operating-point value. That pins the unknown to lb == ub == u0, and NonlinearSolveBase's bounds transform then computes logit((u - lb) / (ub - lb)) = logit(0/0) = NaN.

This fix restores the ±Inf sentinel after resolution for entries whose metadata bound was already ±Inf, leaving the resolution machinery (and the ability of a bound to reference another symbol) intact. The all-numeric branch never entered this path and is unchanged.

Found while investigating #4672. It is not that issue's reported failure — see the analysis in that thread — but it produces the NaN that issue's title describes, via a different route. It only fires when at least one bound is symbolic, which is why the numeric-bounds MWEs in #4672 don't hit it.

Failing before / passing after

Test added to the existing bounds testset in lib/ModelingToolkitBase/test/nonlinearsystem.jl.

Before the fix (test applied to unmodified source at 453c45c08e, source change stashed):

Symbolic bounds ... : Test Failed at testcase.jl:13
  Expression: sprob.lb[sdi] == -Inf && sprob.ub[sdi] == Inf
Symbolic bounds ... : Test Failed at testcase.jl:14
  Expression: SciMLBase.successful_retcode(ssol)
Symbolic bounds ... : Test Failed at testcase.jl:15
  Expression: !(any(isnan, ssol.u))

Test Summary:                       | Pass  Fail  Total   Time
Symbolic bounds ...                 |    1     3      4  26.1s
ERROR: Some tests did not pass: 1 passed, 3 failed, 0 errored, 0 broken.

Observed values on the unfixed code:

lb = [-1.0, 0.4]   ub = [1.0, 0.4]   u0 = [0.8, 0.4]     # 0.4 is the unbounded unknown's guess
solve  ->  retcode = Unstable,  u = [0.8, NaN]

After the fix:

Test Summary:                       | Pass  Total   Time
Symbolic bounds ...                 |    4      4  15.8s

Verification

GROUP=InterfaceII (the group that owns nonlinearsystem.jl), on this branch:

Test Summary:        | Pass  Total     Time
NonlinearSystem Test |  106    106  2m34.4s
...
4315.088507 seconds (3.78 G allocations: 359.776 GiB, 4.11% gc time, 85.43% compilation time)
     Testing ModelingToolkitBase tests passed

Runic: clean on both changed files. typos: clean.

Not verified / known gaps

• GROUP=QA fails on this branch — but it fails identically on unmodified master. I ran QA on a clean 453c45c08e worktree to check: both give 10 passed, 1 failed, both failing the Aqua Piracy check at Aqua/src/piracies.jl:244, with a byte-identical method list (src/ModelingToolkitBase.jl:119,123,131,132,133 and src/problems/jumpproblem.jl:257,263). None of those files are touched here. Pre-existing master failure, reported separately; not introduced by this PR.
• Docs build not run — this PR adds no public API and no rendered-docs entry (restore_infinite_bounds is an internal helper; its docstring is not registered in docs/).
• Other test groups, downstream packages, and GPU paths not run.

Worth pushing back on

• The helper restores the sentinel after resolution rather than seeding the map with ±Inf up front. Seeding would be tidier, but it would change behaviour for a bound that references another unknown: today such a bound resolves through that unknown's operating-point value, and seeding would make it resolve through that unknown's bound instead. I kept the existing semantics rather than silently changing them.

Environment: Julia 1.12.6, ModelingToolkitBase v1.61.0, NonlinearSolve v4.26.0, base commit 453c45c08e.