- Return the correct NaN when an invalid operation generates a NaN.
  - When one or more operands are NaN, handle them as the processor
    would, prioritising signalling NaNs and making them quiet.
  - Fix fmadd and related instructions:
     - Fnmadd is fma(-n, m, -a), not -fma(n, m, a).
     - Some common libc implementations incorrectly implement fma for
       zero results, so work around these cases.
  - Replace some unreliable tests.

This patch also adds support for Default-NaN mode, since once all the
other work was done, it only required a couple of lines of code.
Default-NaN mode was used for an optimisation in ARM, and it should now
be possible to apply the same optimisation to A64.

BUG=
R=jochen@chromium.org

Review URL: https://codereview.chromium.org/199083005

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@19927 ce2b1a6d-e550-0410-aec6-3dcde31c8c00