The Quokka2026 EMF scheme shows instabilities for slow waves at high resolution

**Describe the bug** The `Quokka2026` EMF compute scheme is prone to produce a numerical instability at high resolution with oblique or complex magnetic field geometries. This was first seen in the Orszag-Tang vortex (runs stalled at t ~ 0.75 at 256^2), and has now been seen again in the slow magnetosonic wave at high resolution. The likely cause is that `Quokka2026` reconstructs face-centered HLLD Riemann velocities (a nonlinear quantity) to cell edges, whereas `FelkerStone2017` reconstructs cell-centered velocities that remain well-behaved for linear perturbations. The error with `Quokka2026` diverges sharply from the expected 2nd-order convergence trend at high resolution, where `FelkerStone2017` and `Balsara2025` remain well behaved. The instability is insensitive to CFL reduction. <img width="3329" height="2070" alt="Image" src="https://github.com/user-attachments/assets/acdf517f-6809-4d8d-95c7-4881f302ad99" /> **To Reproduce** 1. Build the `SlowWave` problem (added in PR #1910). 2. Add the following overrides to `inputs/SlowWave.toml`: ```toml amr.n_cell = [512, 8, 8] amr.max_grid_size_x = 128 amr.blocking_factor_x = 16 mhd.emf_compute_scheme = "Quokka2026" mhd.emf_averaging_scheme = "Balsara2025" ``` 3. Run: `./SlowWave ../inputs/SlowWave.toml` 4. See that the L1 error is orders of magnitude larger than the expected converged errorr. **Additional context** - The instability is specific to the `Quokka2026` EMF compute path; it is not present for `FelkerStone2017` or `Balsara2025` at the same resolution. - The EMF averaging scheme (`LondrilloDelZanna2004` vs `Balsara2025`) does not affect whether the instability appears, only the compute scheme does. - Physical resistivity suppresses the instability. This is the proposed fix, to be implemented in a follow-up PR. `