New Kastaun Recovery

[c-prather]

The existing dev branch contains an attempt at applying simulation floors within the Kastaun inverter in a way which prevents positive feedback loops.

However, it's a bit ad-hoc. The remedy for low density wasn't really physical (luckily it was almost never triggered in practice!) and the remedy for low internal energy amounted to stealing as much kinetic energy as necessary to satisfy the floored temperature, thus preserving only the total energy and not the momenta, where one might reasonably hope to do the opposite.

Often, the re-partitioning of energy failed. The primary "fixup" option was just to replace the primitive variables with atmosphere, giving up on preserving any of the "conserved" variables at all. It was crude but undeniably effective, and triggered in maybe 1 in 1e3 grid zones, which I think overall would have been fine.
(The other option, averaging neighbors, is really only effective for more forgiving floors -- once zones with high density bordered empty zones with high velocity, problems ensued quickly and it was never terribly stable).

In the interest of doing yet better, and specifically of being able to document the scheme without humiliating myself, this is take two: same stability, more consistent philosophy. Specifically: trust the inverter. Rather than apply floors during inversion, and then mess with the solution before returning it, this PR simplifies the inverter to essentials, preserving exactly the prescription guaranteed to converge, coupled with only the strictly necessary limits for numerical stability (keeping numbers positive and speeds barely liminal).

Three mitigations are then applied after the inversion:

1. Floors are applied to rho and u in the normal observer frame, just like they are when one sets floors/frame=normal. This application must be in the normal frame, as we still want to conserve momentum if possible (not to mention we only have access to the conserved variables at this point). If the user has selected frame=normal, all floors are applied here -- otherwise, only the minimal floors (rho/u_min_const) are applied here, and the rest are added in the desired frame later, in the usual floor call.

2. After these floors are added to conserved variables, if the inversion afterward cannot reproduce the momenta, reduce the velocities using the old recovery code. That is, solve for $\mu$ from Kastaun's expression for $\hat{\epsilon}$, setting it equal to our known $\epsilon$ after floors. This amounts more or less to prioritizing the thermal and magnetic energy when partitioning the total $T^0_0$, and only using the remainder of the total for kinetic energy and hence velocity. That is, it dissipates kinetic into thermal energy, which is what the code does all the time anyway. I believe it does not actually conserve $T^0_0$, but we already broke conservation of total energy when we applied the floors in normal frame, so it doesn't really matter.

3. If two inversions and a solve for new velocities all failed, set the velocities to zero (alternatively, just fall through to setting all primitive vars to atmosphere -- both strategies look nearly identical).

Ideally, very few zones (fewer than before) should hit (3). The combination of strategies here tries to conserve momentum first, and if this proves impossible, to at least do something plausibly physical before giving up and setting atmosphere.

[c-prather]

Sigma of a million isn't cool. You know what's cool? Sigma of a billion

log_sigma_poloidal.mp4