Eliminate boundary equations to improve HYPRE solves

[rfalgout]

This PR adds code to eliminate boundary equations to get better behavior in HYPRE solvers.

This is not fully tested yet and likely requires some reorganization and cleaning up.

[rfalgout]

Hi @bendudson . I needed the rhs vector in the routine that updates the solution values https://github.com/boutproject/BOUT-dev/blob/5ce6cae6ea94a76417a4c058fdd848db384d36d1/src/invert/laplace/impls/hypre3d/hypre_boundary.c#L211 I couldn't figure out a nice way to do that so I added the following member function to the HypreVector class https://github.com/boutproject/BOUT-dev/blob/5ce6cae6ea94a76417a4c058fdd848db384d36d1/include/bout/hypre_interface.hxx#L241 If you have better ideas for passing the information around, please let me know.

Also, I've made a number of assumptions about the structure of the boundary equations relative to internal equations. See the comments at the top of the file in the first link above. One thing I don't think I mention in this file is the assumption that internal equations that couple to a boundary equation are assumed to live on the same processor as that boundary equation.

[ZedThree]

Unfortunately this PR has broken some of the Hypre tests:

The integrated test test-laplace3d-hypre:

Making Laplace 3D with Hypre
running on 1 processors: ./test-laplace3d -d data_slab_core
 => data_slab_core passed with maximum error 6.382497419465949e-09
running on 1 processors: ./test-laplace3d -d data_slab_sol
 => data_slab_sol passed with maximum error 2.7718056405667824e-08
running on 1 processors: ./test-laplace3d -d data_circular_core
 => data_circular_core passed with maximum error 1.478270839072593e-10
running on 1 processors: ./test-laplace3d -d data_circular_core-sol
 => data_circular_core-sol failed with maximum error 0.00034150972873735197
=> Some tests failed

And the unit test (just showing the first few points, because it's very noisy):

/home/runner/work/BOUT-dev/BOUT-dev/tests/unit/invert/laplace/test_laplace_hypre3d.cxx:165: Failure
The difference between expected[i] and actual[i] is 2.880420150274329e-06, which exceeds tol, where
expected[i] evaluates to 2.4848271042405023e-06,
actual[i] evaluates to -3.9559304603382663e-07, and
tol evaluates to 1e-08.

/home/runner/work/BOUT-dev/BOUT-dev/tests/unit/invert/laplace/test_laplace_hypre3d.cxx:165: Failure
The difference between expected[i] and actual[i] is 0.025529941378404492, which exceeds tol, where
expected[i] evaluates to 0.021339352519864602,
actual[i] evaluates to -0.0041905888585398914, and
tol evaluates to 1e-08.

/home/runner/work/BOUT-dev/BOUT-dev/tests/unit/invert/laplace/test_laplace_hypre3d.cxx:165: Failure
The difference between expected[i] and actual[i] is 2.2593733713480629e-05, which exceeds tol, where
expected[i] evaluates to 2.0087612157650606e-05,
actual[i] evaluates to -2.5061215558300253e-06, and
tol evaluates to 1e-08.

/home/runner/work/BOUT-dev/BOUT-dev/tests/unit/invert/laplace/test_laplace_hypre3d.cxx:165: Failure
The difference between expected[i] and actual[i] is 0.42204141661955613, which exceeds tol, where
expected[i] evaluates to 0.38156718593235334,
actual[i] evaluates to -0.040474230687202815, and
tol evaluates to 1e-08.

/home/runner/work/BOUT-dev/BOUT-dev/tests/unit/invert/laplace/test_laplace_hypre3d.cxx:165: Failure
The difference between expected[i] and actual[i] is 3.0573430513351925, which exceeds tol, where
expected[i] evaluates to 2.9010150915766806,
actual[i] evaluates to -0.15632795975851177, and
tol evaluates to 1e-08.

This is in the first test TEST_P(LaplaceHypre3dTest, TestMatrixConstruction3D), and then gtest crashes completely.

@rfalgout This test compares the inverse to the forward solution. It's possible our forward operator in the test has some issues at the boundaries

@bendudson It might be wise to back out these changes until we get this sorted?