On Using Floating-Point Computations to Help an Exact Linear Arithmetic Decision Procedure