Investigate using symmetric decomposition for the implicit method

MUMPS allows to use LL^T decomposition, which appears to give a 50%–100% speedup to the computation and should save half the decomposition memory usage (which is the largest memory cost in the implicit method).

Switching the decomposition is as simple as using ctx.set_matrix(mat, symmetric=True), however it also breaks the tests, and will likely require investigating what is up.