Add an L-BFGS rule with a ring-buffered memory and an mlx dispatch

[jessegrabowski]

Adds L-BFGS at a fixed step size. lbfgs_updates in rules.py keeps per-parameter stacks of the last memory_size parameter and gradient differences as a ring, written in place with set_subtensor, and admits a pair only when y . s > eps * y . y, so the inverse-Hessian estimate stays positive definite. LBFGSDirection, a SymbolicOp in optim/lbfgs.py, applies the two-loop recursion to the gradient with two scans over the ring order and gets a Python-loop dispatch in dispatch/mlx/lbfgs.py, since Scan has no mlx dispatch. The lbfgs alias takes a rate or a schedule like the other rules.

The ring slot is a one-element index vector rather than a scalar. pytensor's mlx Subtensor dispatch cannot trace a scalar index (pymc-devs/pytensor#2422), and the advanced-indexing ops can.

The mlx CI install now requires mlx>=0.32.2. Below that, mlx's vector matmul ran one threadgroup and a 12.6M-element dot took 155 ms instead of 0.4 ms (ml-explore/mlx#3580).

state_for takes history_size for the stacked buffers and scalar_state takes a dtype for the pair count. Tests check the op against the dense BFGS matrix built from its definition and the rule against the secant condition and a quadratic's closed-form minimizer, on numba and on mlx.

The line search is the next PR. Part of #58.

---

📚 Documentation preview 📚: https://pytensor-ml--162.org.readthedocs.build/en/162/