Document that LCAO will not work with combined k and state parallelisation

LCAO initial guess does not work with combined k and state parallelisation. It looks like this got lost in the docs between V15 and 16 (due to moving to a single LCAO implementation?). Additionally, the code was terminating with an ASSERT. I've changed this to a fatal message, as this is due to conflicting input options rather than a dev calling the routine erroneously.