Fix energy conservation with sparse autodiff

[dennisYatunin]

This PR fixes energy conservation when using AutoSparseJacobian by adding padding bands to most of the ∂ᶜρe_totₜ/∂ᶜχ blocks, and also by explicitly handling quaddiagonal bands of the ∂ᶜρe_totₜ/∂ᶠu₃ block.

Based on the dry and moist Float64 baroclinic wave conservation tests, errors in the ∂ᶜρe_totₜ/∂ᶠu₃ block as small as one part in a trillion seem to affect energy conservation. The new bands eliminate all such errors, so that energy conservation is even better than with ManualSparseJacobian. Further experiments may show that this is more precise than we need, so we could potentially improve performance by removing some of these new bands. For now, though, we can continue testing sparse autodiff with optimal energy conservation.

This PR also adds a dry Float64 baroclinic wave conservation test with sparse autodiff to CI, removes the soft-fail annotation from the moist test, and clarifies the bandwidth tables printed when debugging the Jacobian.

---

• I have read and checked the items on the review checklist.