Avoid hitting fp16 NaN values

Avoid hitting fp16 NaN values

Modify the test references values to avoid situations where
they hit NaN values when viewed as fp16. This can happen
when uploading {0.1, 0.5, 0.9, 0.2} to r8g8b8a8_unorm which,
depending on the rounding of the 0.5, could end up as a 0x7F
value which when viewed as the MSB of a fp16 value is a NaN.

Affects:

dEQP-VK.image.mutable.*

Components: Vulkan

VK-GL-CTS Issue: 784

Change-Id: I2adbd27aa7d94254c9f19d788eacace2ea95d10c
(cherry picked from commit d51c248f04b980406d9b767064170b77b682cac7)