This patch fixes the numerical errors in aarch64 sqrt() due to the
compilers use of fnmsub (negated fused multiply-subtract).

Previous thread:
https://sourceware.org/ml/libc-alpha/2013-12/msg00473.html

I'll leave this on the list for 24 hours before committing to give
folks the opportunity to comment on the (in)appropriateness of my
solution...

Cheers
/Marcus