Search every L-BFGS step with a zoom line search by default

[jessegrabowski]

lbfgs_updates, the L-BFGS rule, now takes an optional line_search. With one set, it searches along the step -learning_rate * H g and moves by the step size the search accepts, so learning_rate becomes the first trial. The lbfgs alias defaults to zoom_line_search(), and on Rosenbrock it reproduces the first 12 losses of optax.lbfgs() to 1e-10.

With the search on, lbfgs() raises when it is given precomputed gradients, when it is placed after another transform in a chain, or when the loss draws random numbers. The search evaluates the loss at several trial points, so it needs the loss itself, and the loss has to be deterministic. line_search=None restores the fixed step.

On mlx, every search runs all max_steps trials, because mx.compile cannot stop a loop early. Each step there costs that many loss and gradient evaluations.

LineSearch and zoom_line_search are now exported from pytensor_ml.optim.

Closes #58

---

📚 Documentation preview 📚: https://pytensor-ml--170.org.readthedocs.build/en/170/