Revert AutoSpecialize default for SciMLFunction constructors

[ChrisRackauckas-Claude]

Summary

• Reverts only the ODEFunction, SplitFunction, and DynamicalODEFunction convenience constructors back to FullSpecialize (from Use DEFAULT_SPECIALIZATION for all SciMLFunction constructors #1298)
• All other function types (NonlinearFunction, SDEFunction, DiscreteFunction, DAEFunction, RODEFunction, etc.) keep DEFAULT_SPECIALIZATION as intended
• Bumps version to 2.153.1

The ODEFunction change exposed multiple breaking issues in OrdinaryDiffEq.jl CI:

1. BigFloat tableau c promotes t past FunctionWrapper signatures — requires tableau constructors to take separate state/time types
2. GradientTracer type not in wrapper signatures — AutoSparse sparsity detection passes tracer types through the wrapped function
3. MatrixOperator jac_prototype causes promote_op → Union{} — promote_f can't handle operator-based Jacobians
4. Hard DAE numerical accuracy regression — wrapping changes numerical path for DAE + callback problems

Deferring ODEFunction AutoSpecialize default to v4 where these can be addressed properly.

Test plan

• OrdinaryDiffEq.jl CI should return to green once this is released
• Other function types remain on DEFAULT_SPECIALIZATION and are unaffected