Use Dirichlet boundary conditions in 2D heat eq problem

[charleskawczynski]

The convergence tests have been flakey for a while, and I think I've found the issue.

The 2D heat equation problem is doubly periodic, but I don't think that we're explicitly enforcing the compatibility constraint. The constraint states that the descrete integral of the source terms must be exactly zero (i.e., within machine precision zero). This effectively removes the null space, and chooses a particular solution that we're interested in.

My rough explanation of this is:

• We're solving an equation for T on a 2D domain, with purely periodic boundaries.
• This means that T + 14 will also satisfy the equations
• If we were to solve this with a matrix-based approach, AT = b, then the condition number of A would be Inf.
• We're not forming a matrix in ClimaCore, we're using an iterative matrix-free method. The result is that the solution is highly sensitive and can easily blow up

There's a few ways to do this:

• In a matrix-based approach, we could subtract the weighted mean of the right-hand-side (RHS) from the RHS. The result would (need to) be that the discrete integral of the RHS is zero. This is an elegant approach, but requires quite a bit of care and extra machinery.
• "pin" the solution by setting the solution anywhere in the domain to be a fixed value

Since this problem uses sin as a source term, we have one more option: use Dirichlet boundary conditions. In this approach, we can set the tendency to zero at the boundary, which also effectively pins the solution.

I think the result is that we should see the solution unchanged, but also now stable.

I'm going to trigger CI a bunch of times and see if this fixes the flaky jobs.

(hopefully) closes #283.

[charleskawczynski]

All 2D heat eq passes:

• build (only one jet job failed, but I've since fixed that)

[charleskawczynski]

Interesting, SSP22Heuns broke. Let's also print debug info from #338.