This reverts commit 33672520

The check fails in recent CI, but I think it should be correct
anyway. Also, I dont think that error bounds below single precision
accuracy are good for numdiffs.

There is way more reasons to interpolate a function into the
solution vector and we cover them with one code path:
* Dirichlet boundary conditions (strong ones)
* Initial conditions for instationary expressions
* Initial guesses for Newton solvers