Fix OptimizationManopt reporting MaxIters even when converged

[SebastianM-C]

Summary

• __map_optimizer_args! fully replaced Manopt's own per-solver default stopping_criterion with just StopAfterIteration/StopAfter whenever maxiters/maxtime was given without an explicit abstol. Since those criteria never indicate convergence (Manopt's has_converged semantics), this made ReturnCode.MaxIters structurally unavoidable regardless of whether the solve had actually converged. Fixes the underlying issue behind Improve return codes from OptimizationManopt #1034, which Improve return codes from OptimizationManopt #1169's retcode-mapping fix exposed but never actually resolved.
• Added a per-optimizer fallback tolerance (_default_fallback_abstol), used only when maxiters/maxtime is set without abstol, mirroring each solver's own real Manopt default so genuine convergence is still reported as Success. Optimizers whose real default is a composite criterion (NelderMead, CMAES, ConvexBundle) opt out and keep prior behavior.
• Fixed the Karcher-mean docs example, which claimed to use Frank-Wolfe with a closed-form subsolver but never reassigned opt from the earlier GradientDescentOptimizer — Manopt silently warned and dropped the unrecognized sub_problem kwarg, so it silently ran plain gradient descent instead.
• Adjusted for the Manopt 0.6 sub_problem calling-convention change (now needs evaluation = Manopt.InplaceEvaluation() for the mutating closed-form style) and bumped maxiters budgets across all three docs examples to values confirmed sufficient for them to reach Success.
• Verified with Manopt.check_gradient that the two examples without an explicit abstol (Rosenbrock/Euclidean and Rayleigh quotient/Sphere) have correct gradients — their MaxIters was genuinely insufficient default iteration budget, not a gradient bug.

Test plan

• Pkg.test() in lib/OptimizationManopt — all 16 tests pass, including the pre-existing Quasi-Newton/ARC/TrustRegions tests
• Added a regression test asserting successful_retcode for a maxiters-only (no abstol) gradient-descent run
• Ran all three docs/src/optimization_packages/manopt.md examples end-to-end; each now reports retcode = Success with the expected minimizer/objective

🤖 Generated with Claude Code