On using an inexact floating-point LP solver for deciding linear arithmetic in an SMT solver