Properly figure out NaN and -0 cases even for non-integer inputs. This
helps to reduce the number of checks we have to perform in case we try
to go back to int32 after a floating point operation.

R=epertoso@chromium.org

Review-Url: https://codereview.chromium.org/2221863002
Cr-Commit-Position: refs/heads/master@{#38415}