Account for differences in nan with fp16 comparison

Account for differences in nan with fp16 comparison

When comparing fp16 values it's not possible to use a bitwise
comparison because there can be many different bit patterns
representing nan values, and all nans are equal. This CL updates
the computeCheckBuffersFloats function which compares two
buffers of fp16 values to take into account different forms
of nan.

Affects:

dEQP-VK.spirv_assembly.instruction.compute.16bit_storage.uniform_16_to_16.stress_test

Components: Vulkan

VK-GL-CTS issue: 1018

Change-Id: I77013b49ce070d9615b287eae24dd32731b314b1
(cherry picked from commit 92d5909ab99bc97dc2ae4dd67e1cfa592dafb246)