Fix OptimizationManopt Hessian on matrix-valued manifolds

[SebastianM-C]

Fixes #1036.

Problem

TrustRegionsOptimizer and AdaptiveRegularizationCubicOptimizer threw a DimensionMismatch on any manifold whose points are matrices (Stiefel, SymmetricPositiveDefinite, ...). build_hessF allocated the Euclidean gradient and Hessian-vector buffers as flat length(θ) vectors, which breaks the projection inside ManifoldDiff.riemannian_Hessian!. The Euclidean tests carrying the TODO from #1009 pass and produce the correct Hessian-vector product, so the failure only shows once points stop being vectors, which nothing in the test suite exercised.

Changes

• build_hessF allocates with zero(θ) so buffers keep the point's shape.
• When no hv is available the Hessian-vector product is assembled from a dense user-supplied hess; with no second-order information at all it returns nothing, so Manopt uses its own approximate Hessian instead of a closure that would throw on nothing.
• The stale test TODOs are replaced by a Stiefel regression testset covering AD, hess-only and grad-only, plus a check of the converted Hessian against riemannian_Hessian with exact Euclidean inputs.
• The Manopt docs gain a section on how Euclidean derivatives are converted and where second-order information is taken from.

Not addressed: requireshessian still makes OptimizationBase prepare a dense Hessian that Manopt never evaluates. That needs an allowshv-style trait in SciMLBase and is left for a follow-up.

🤖 Generated with Claude Code