Fix subnormal floats in extended image usage tests
When transcoding images in the extended image usage bit tests, Inf and
NaN were correctly taken into account and removed from input data.
However, subnormal floats were only being taken into account in the
32-bit float case.
The Vulkan spec regarding the precision and operation of SPIR-V
instructions mentions subnormal values may be flushed to zero in any
shader instruction, including the shader code to transcode images, and
have to be taken into account for half floats too.
Afffected tests:
dEQP-VK.image.extended_usage_bit.*
Components: Vulkan
VK-GL-CTS issue: 1747
Change-Id: I74a5d27b40f773f8d85e86a0a5d04eb0b4f16194
(cherry picked from commit
300b27043f8c93dcce4b3c788b33688ab59867af)