Use the analytic event Jacobian as the AD test oracle

[ChrisRackauckas-Claude]

Ignore this PR until it has been reviewed by @ChrisRackauckas.

Summary

• replace the bouncing-ball event test's finite-difference Jacobian oracle with its exact one-bounce Jacobian
• use the same analytic reference for Jacobian and gradient backends
• leave every tolerance and the existing broken Enzyme assertion unchanged

This is a test-oracle correction; it does not change solver behavior.

Clean-master failure

The trajectory contains exactly one bounce. Perturbing the parameters for finite differences also perturbs the located root time, so the finite-difference reference carries enough event-time noise to fail the default comparison:

• ForwardDiff differs from the analytic Jacobian by at most about 2.7e-9
• the finite-difference oracle differs by about 7.5e-8

On untouched master a7080c6f6d, the full AD group passes its 28,299 general checks, then the event set ends with 3 passes, 2 failures, and 1 existing broken test. The failures are the ForwardDiff Jacobian and Enzyme gradient comparisons against that noisy finite-difference oracle.

Hosted history brackets the exposure after the last passing 275ed1e4 run and the shared-controller merge fd3f95bb99 (#3720); the test itself did not change across that boundary. The closed-form reference distinguishes a small numerical trajectory shift from an AD regression.

Analytic reference

For initial height h and restitution e, the impact speed is v = sqrt(2h). The final-state Jacobian used here is:

[(1 + e) / v - 1/2    v - 2
 (1 + e) / v - 1      v]

Local verification

At 05db644879, the exact full GROUP=AD julia +1.12 --project -e 'using Pkg; Pkg.test()' run passes:

• AD tests: 28,299/28,299
• autodiff event tests: 5 passes and the same 1 existing broken test
• discrete adjoint: 1/1
• Runic on the changed file: pass
• git diff --check: pass