Add WolfeSearch: a strong Wolfe line search

[MashedP]

Summary

Adds WolfeSearch, a line search satisfying the strong Wolfe conditions, and
makes it selectable on the quasi-Newton minimisers. This addresses the
# TODO(raderj): switch out BacktrackingArmijo with a better line search
markers in BFGS/DFP: a Wolfe search enforces a curvature condition, which
keeps steps large enough for quasi-Newton Hessian updates to stay
well-conditioned, and in practice converges in fewer iterations.

For example, on Rosenbrock from [-1.2, 1.0], LBFGS with the default
BacktrackingArmijo stalls, while LBFGS(search=WolfeSearch()) converges to
[1, 1].

Design

WolfeSearch follows the bracketing/zoom algorithm of Nocedal & Wright
(Alg. 3.5,3.6), using bisection for the zoom interpolation. Because Optimistix
calls AbstractSearch.step once per solver step (the iteration loop belongs to
the solver), the textbook nested loops are expressed as a single per-call state
machine dispatched by a mode flag, no nested lax.while_loop.

The curvature condition needs the gradient at each trial point, which the
solvers currently do not pass to the search. To avoid penalising searches that
only need function values (e.g. BacktrackingArmijo), this adds an opt-in:

• AbstractSearch.requires_grad_eval: ClassVar[bool] = False.
• WolfeSearch sets it to True.
• AbstractGradientDescent.step and AbstractQuasiNewton.step compute the
  trial-point gradient eagerly only when search.requires_grad_eval, and
  reuse it on acceptance (so there is no extra work overall, and no change for
  the default path).

BFGS, DFP and LBFGS gain an optional search= argument; the default
remains BacktrackingArmijo, so existing behaviour is unchanged.

Tests

• New tests/test_wolfe_search.py: verifies the strong Wolfe conditions hold
  directly on a 1D restriction, convergence of BFGS/LBFGS × use_inverse with
  Wolfe, jit, the requires_grad_eval flag, and that the default Armijo path
  is unchanged.
• All 493 existing tests/test_minimise.py cases still pass.
• ruff clean; beartype-strict typechecking passes.