[gpu-3] Share polar spline factorizations across solvers

[max-models]

Building multiple polar projector and mass-preconditioner instances repeatedly factorizes identical matrices and creates duplicate device copies. This change shares live 1D and poloidal factors across the polar spaces of one Derham object.

Stack: based on gpu-2-vlasov-benchmark (#735), as the next PR in the GPU stack.

The cache matches exact canonical sparse matrix contents and dtype, solver kind and ILU settings, Schur pole masks/outer factors, and backend/device. It reuses both host factorizations and their cached GPU solves. Boundary-condition and transposed variants share only when their matrices and configurations actually match. Updated mass weights obtain new poloidal factors while retaining unchanged 1D factors.

Entries use weak references: live solvers keep their factors, while obsolete factors and their keys are released when no solver uses them. Different Derham objects have separate caches.

In the small-cylinder fixture, constructing all four solve variants for P0–P3 reduces sparse LU calls from 88 to 9, and dense Schur LU calls from 24 to 8.

Validation:

• Cache tests cover sparse storage normalization, exact value/dtype matching, solver settings, Schur structure, transpose separation, mass-weight updates, lifetime cleanup and isolation between Derham objects/backends.
• Serial polar/cache/backend regression suite: 57 passed, 19 skipped.
• Two MPI ranks, polar and cache tests: 50 passed, 1 skipped per rank.
• NVIDIA T1000, CuPy 14.2.0: 7 GPU cases passed, including both mappings and all three solvers; shared-device-factor reuse also passes after the final ownership change.
• Reusing warmed GPU factors introduces no cunumpy host/device transfers.
• Ruff lint/formatting, git diff --check, and commit hooks passed. Assembly kernels compiled with GNU/Fortran for local verification.