Compute linear residual including Lagrange restrictions

[pjaap]

No description provided.

[chmerdon]

Besides this really weird error on the macOS test of Example212 I noticed another problem (unrelated to the changes in this PR): when I want to solve a problem with two restrictions with constant_matrix = true for different right-hand sides, the second run crashes, because the QR decomposition is not stored and therefore the right-hand sides of the second run is larger than in the first run, while the matrix stays the same. To fix this, I guess we either need to store the QR decomposition or enforce that constant_matrix cannot be true when compress_restrictions = true. I think I am in favor of the first option. But let's first try to understand/fix the other issue.

[chmerdon]

I am unsure about uniqueness in this example. I think the mean value of the displacement iin the periodic x direction should vanish. We could enforce this by adding another restriction

assign_restriction!(PD, ZeroMeanValueRestriction(u; operator = ExtendableFEMBase.IdentityComponent{1}))

The result looks similar, but the test fails. Maybe we can check if all three architectures agree in this case and redefine the test?

[pjaap]

Merging now, since the other problem are unrelated to this PR. The failure for macOS also occurs on master. That and the problem with constant_matrix are attacked separately.