Fix discrete adjoint when there are extra time steps #426
Description
It seems that there is an issue in the discrete adjoint when we use extra time steps as what is needed to compute the adjoint.
Surprisingly the discrete adjoint has a higher error when we use LossH and tV_ref isn't empty. A store argument was added in #420 to mitigate this so that (i) we don't compute the solution where it's not needed and (ii) the discrete adjoint has a better accuracy when we use only LossH (see this thread #420 (comment)).
Obviously this doesn't fix the underlying bug and we should address it properly