Fix AMR with the new energy model

[marchdf]

The key line is increasing the m_f_nghost from 2 to 3

[marchdf]

This makes is so truly adaptive mesh refinement works. The sod_amr test case has a tagging based on density and the refinement zone follows the shock:

Green: level 0 data, red: sim with an extra level at the shock. You can see that the shock is sharper.

[marchdf]

The only issue right now is that the outflow BC doesn't work. So if you use this as is, it will try to refine the outlet and then fail. I had to switch the outlet to foextrap to have this run. So I am going hit pause on this and fix that first.

[marchdf]

ok fixed the outflow BC issue thingy by ensuring there are valid cells (foextrap) in the domain boundary cells when getting a field for tagging.

[marchdf]

And that last commit fixes things when there is mesh refinement at the BC:

Because:

• level 1 takes 2 steps for level 0's 1 step, and because
• our BC applier only applies a physical BC on the distributions pointing directly into the domain (otherwise it fills with -1)

Therefore:

• we need to apply the physical BCs at each substep of a level advance. This ensures that the distribution pointing inwards is always valid. If we didn't do this, then on the second step of level 1, we would have a -1 from the second ghost cell (having reached the first ghost cell after the first step), come into the domain. Reapplying the physical BCs ensures this never happens.