Make supplied linearisation reuse deterministic

[nicamo]

Closes #92.

Motivation

Issue #92 identified that a bare cached dIydI does not record the passive-mode basis used to build it. Coupling-based selection and post-Jacobian mode removal can change that basis, making reuse brittle or dimensionally inconsistent.

Changes

• Accept a supplied dIydI only with mode_selection="timescale".
• Reject coupling-based reuse immediately, including when mode_removal=False.
• Apply the same validation to constructor input and later initialize_from_ICs(..., dIydI=...) calls.
• Validate supplied dIydI and dIydtheta dimensions against the selected basis, grid, and profile parameterisation.
• Copy supplied arrays so subsequent solver operations cannot mutate the caller's cached matrices.
• Document the compatibility contract in the solver API.
• Add executed Example 05d showing how to cache and reuse a linearisation, run a GS-backed linear step, and recognize unsupported configurations.

Scope

This does not remove mode_removal. Fresh coupling-based linearisations can continue to use it. Timescale selection continues to disable post-Jacobian mode removal automatically. The restriction applies specifically to externally supplied current-response Jacobians, whose column basis must be deterministic.

Validation

• freegsnke/tests/test_mode_selection.py: 6 passed.
• Example 05d executed end to end with six linearisation workers and no error outputs.
• The reused and source solvers produced matching cached matrices, current updates, and GS-backed jtor after a linear evolution step.