Correct for the odd offsetting that math/bigfloat does for exponents

[pavpanchekha]

This PR fixes failures in FPBench. Basically, in Racket's math/bigfloat package, the "exponent" of a bigfloat isn't what we normally think of as the exponent; it's more the exponent of 1 ULP. We gotta correct for that by adding (bigfloat-precision x).

[bksaiki]

Good find. The fix may be better written as:

(define e (sub1 (+ exp (bigfloat-precision x))))
(define emin (mpfr-get-emin))
(cond
   [(< e emin) ; subnormal
...

But I think this works.

[pavpanchekha]

Ah, sorry, I was going to pretty up the code after this meeting. Whatever.