Second-order AMR: higher-order field refinement + coarse–fine temporal fixes

[UCaromel]

Why

AMR on master is not second-order accurate. The field refinement (coarse→fine spatial interpolation) and the coarse–fine time interpolation are only first-order, so an AMR run does not recover the scheme's second-order accuracy near refinement boundaries — accuracy is effectively capped below 2nd order wherever levels exist.

This PR makes AMR genuinely second-order in both space and time:

• second-order spatial field refinement, and
• second-order coarse–fine temporal interpolation — including for B, which previously had none.

The existing first-order (legacy) refinement path is kept. Flipping the default and/or removing the legacy path is a team decision; for now both coexist and are selectable at runtime, so this PR changes no default behaviour.

How

• Kernel-based refinement operator — spatial refinement is expressed as composable stencil kernels rather than a hardcoded scheme. This is what makes 2nd order clean, and makes it easy to extend to higher spatial orders and to swap in a WENO-limited variant.
• Correct stage times — RK integrators and the PPC predictor/corrector fill coarse–fine ghosts at the proper sub-step times rather than at endpoint values.
• Time interpolation everywhere — coarse–fine temporal interpolation now covers all evolved fields, including B (previously static in time).

Selectable from the Python config via the refinement order (and limiter); legacy = order 0.

Validation

2D AMR spatial convergence (composite L2-of-L1 norm, MC2011). With refinement_order=2, both alfven2d and whistler2d converge at slope 2, inside the calibrated gate band (1.70, 2.25):