investigate numerical discrepancies btween fortran and torch for Kubi's ecosmo port

they almost match but some issues seem to persist.

could help to closely examine the first few time steps, or to zero some parameters or concentrations to simplify the scenario and identify causes.