Change method used by EXACT->INEXACT on integers. It turns out that

Change method used by EXACT->INEXACT on integers.  It turns out that
the old method, the INT:->FLONUM, does not round reliably, and as a
consequence the LSB of the result is sometimes wrong.  However, the
conversion performed by INTEGER->FLONUM is accurate provided that the
integer being converted can be exactly represented by a flonum, i.e.
for IEEE double-precision floats, an integer with magnitude less than
(EXPT 2 53).

The algorithm used to convert ratnums to flonums already has this
property, so the integer conversion has been changed to use it.