box local residual: choose epsilon more intelligently

we now take the geometric mean of the numerical precision of the
underlying floating point type and 1. for standard 64 bit double
values this does not change anything (the base epsilon stays 1e-8),
but for higher or lower resolution scalars, the value is now "always
right". probably.

git-svn-id: svn://svn.iws.uni-stuttgart.de/DUMUX/dumux/trunk@6550 2fb0f335-1f38-0410-981e-8018bf24f1b0