Use std::numeric_limits<double>::quiet_NaN() and
std::numeric_limits<float>::quiet_NaN() instead.

Review URL: https://codereview.chromium.org/864803002

Cr-Commit-Position: refs/heads/master@{#26195}