MSVC's implementation of numeric_limits currently generates incorrect
signaling NaNs. The resulting values are actually quiet NaNs instead.
This commit is based off of a solution by shuffle2. The only
difference is a template specialization for floats is also added
to cover all bases