Add FlashMD energy rescaling to fix metatomic

[johannes-spies]

Implements the energy-conservation enforcement from App. C of the FlashMD paper (arXiv:2505.19350): with rescale_energy on, the predicted momenta are rescaled after each step by alpha = sqrt(1 - (E'-E)/K') so the total energy matches its pre-step value.

The potential energy U is taken from a pair_style metatomic energy model on top of the fix, read every step via an internal compute pe (mirroring how fix npt obtains the virial). The rescale is applied in post_force, i.e. after the FlashMD map and before the thermostat half-step, matching i-PI's operator splitting.

Verified on a 108-atom fcc Al NVE example: total energy drifts ~30 meV/atom without rescaling and is flat to machine precision with it, consistent with an i-PI reference and with an independent recomputation of the energy from dumped frames. Example inputs added under examples/PACKAGES/metatomic.

fix metatomic/kk overrides initial_integrate/post_force, so the host-side rescale does not run there; it now errors out instead of silently doing nothing. On-device Kokkos support is left as follow-up.

[Luthaf]

I don't love doing this inside the engine. Would it make sense/be possible to implement this by having the FlashMD model request the energy as an extra input, and do the rescaling internally?

[johannes-spies]

Hm, can the models have state/remember the last energy? Because that's what's needed here.