WIP: Better preconditionner for stack_solve_algo={2,3} (GMRES + BiCGStab solvers)

On first iteration, do a full LU decomposition, and use it as a preconditionner
in further iterations (similar to UMFITER in TROLL).

TODO:
– better understand how the tolerance of GMRES and BiCGStab influence the
  result (hardcoded to 1e-12, needed for GIMF in the block-decomposed version)
– do the same in bytecode
– leave ILU as an alternative preconditionner?
– handle the error flag of gmres and bicgstab?
– remove the persistent variables in sim1.m

[skip ci]