Change the initialization logic for `useCollograd`.

[YohannDudouit]

Avoid using the collocated gradient algorithm when we don't compute any gradient.

[YohannDudouit]

This same change probably also goes in backends/hip-gen, right?

Yes, if you're fine with the change I can spend some time doing it in hip-gen, it should be quick. @nbeams any objection?

[YohannDudouit]

This same change probably also goes in backends/hip-gen, right?

@jeremylt Change propagated to hip-gen.

[jeremylt]

Thanks - we can merge once CI does its thing

[jeremylt]

This change seems to reintroduce the bug we saw in the fluids example, according to the CI

[jeremylt]

CI got suck, but the PETSc tests still seem to be failing.

[jeremylt]

The CUDA CI flaked, but the HIP CI is happy now.

[YohannDudouit]

The CUDA CI flaked, but the HIP CI is happy now.

Should I run make test on Lassen?

[nbeams]

Does the setup on Lassen do the PETSc tests?

[jeremylt]

This has identical changes in CUDA and HIP - CUDA CI is borked but HIP CI is passing so I think we are ok to merge

[YohannDudouit]

This has identical changes in CUDA and HIP - CUDA CI is borked but HIP CI is passing so I think we are ok to merge

@jeremylt What would I need to do to reproduce the CI tests locally?

[jedbrown]

I merged from main and kicked off a pipeline. Assuming it's good, I think we can squash-merge.