Use float16 precision for testing mediump

Use float16 precision for testing mediump

The precision requirements for mediump are not in the spec, they're just
for the conformance test, so it makes sense to use the simplest thing.
Allow the same error margins for mediump's 16-bit as for true 16-bit
floats. This slightly increases the precision requirements on mediump
exp but loosens them for atan.

Note, however, that the requirements for fp16 are still stricter than
for mediump because of the FloatFormat involved. fp16 requires support
for a wider exponent range than mediump.

Components: Vulkan
Affects: dEQP-VK.glsl.builtin.precision.exp*.mediump.*
         dEQP-VK.glsl.builtin.precision.atan*.mediump.*

Change-Id: I600a53984ba6a4f624bbb1525636f392d74ae96f
(cherry picked from commit b7dea11e6a395552a1423009c33d5518d9a0b2ec)