Correctly handle normalized formats in FloatFormat
For the texture explict lod tests normalized values were treated as 8
bit floating point formats when calculating the reference interval.
This ignores the fact that normalized formats are encoded as 1 / N,
where N is an integer, this results in the rounding used in tests not
being able to accurately round normalized formats.
This incorrect rounding causes the acceptable ranges to be
significantly smaller than other formats and extremely asymetric
around the ideal filtered value for certain samples.
Affects: dEQP-VK.texture.explicit_lod.*
Components: Vulkan
VK-GL-CTS issue: 2189
Change-Id: Ia4f3adefab4bfb5d9c91bd7033199b0e107179d5