Use zero instead of similar to avoid undefined behavior

[charleskawczynski]

This will help prevent #283 and #301 from resurfacing for users.

[charleskawczynski]

The bad news is that I don't think that this PR is fixing the issue. The good news is that I did see NaNs when running the unit tests locally. It's surprising, but then again, I don't think that we actually have tests for them, and they're not failing like the docs because we don't plot in the unit tests.

I hope that this can accelerate the bug hunting.

[charleskawczynski]

Ok, good news, I'm pretty sure that in T_imp! was the issue after all. We unfortunately can't (yet) use Base.zero for those input arrays because ClimaCore.MatrixFields.FieldMatrixWithSolver does not support Base.zero, nor does it support Base.fill!. Let's initialize the tendency for now, update ClimaCore, and then we can change CTS to require Base.zero / Base.fill! (for the next breaking release).

[Sbozzolo]

Could you please run a ClimaAtmos test suite to verify that everything still works there?

I think that this change will also implicitely raise the lower compat for ClimaCore in all our packages (to the version that contains the definion of zero).

[charleskawczynski]

It works 🙂

[Sbozzolo]

Thank you!