Use RuntimeGeneratedFunctions instead of eval

[dpsanders]

Summary

• Replaces the two runtime eval(...) calls that build closures (in gradient and forward_backward_contractor) with @RuntimeGeneratedFunction, so callers can build and invoke a generated function in the same dynamic extent without hitting Julia world-age errors.
• Adds RuntimeGeneratedFunctions 0.5 as a direct dependency and initializes it in the module.
• Load-time @evals in reverse_icp.jl (per-operator _rev_* wrappers registered with Symbolics) are left alone — they don't suffer from world-age issues.

Notes

• RGF doesn't accept destructuring patterns in argument position, so the generated closures now take plain argument symbols (__args, __inputs, __constraint, __params) and destructure on the first line of the body. Calling convention from the user's side is unchanged.
• Stacked on top of Add SSAFunction type, refactor pipeline to accept SSA input #80 (add-ssa-function).

Benchmarks (median; see commit message / discussion for full numbers)

Operation	eval (baseline)	RGF
gradient build, 2 vars	907 μs	126 μs (7.2×)
gradient build, 3 vars + transcendentals	2.08 ms	555 μs (3.8×)
forward_backward_contractor build, x²+y²	1.39 ms	89 μs (15.6×)
fwd_bwd_contractor call	479 ns	504 ns (noise)

Steady-state call cost is essentially unchanged in realistic use.

Test plan

• Pkg.test() passes locally (12/12)
• Smoke test: build and invoke a generated gradient and contractor inside a single function body (previously would raise a world-age error)

🤖 Generated with Claude Code