Fix bug in unsigned modular arithmetic using a signed implementation

Fix bug in unsigned modular arithmetic using a signed implementation

If we aim to be clever by implementing an unsigned modular arithmetic
computation using signed arithmetic, we need to make sure that we
don't accidentally contaminate the computation with any extraneous
high bits.  This means that we must be sure to cut constants to the
appropriate width, as well as computations, so do so; this fixes
bug #974406 from Paul Dietz.  (In addition the change from cutting
to the requested width to the implementation width fixes #903821,
so Go Team!)

Test cases.  Minimally horrible test case for #903821; far worse
suggestions were made on #sbcl IRC...