Add support for float colours on the GPU

[DJMcNab]

As requested by @LaurenzV in #967.

This probably shouldn't be merged. Note that a tolerance of zero still doesn't work, for unknown reasons.

One potential reason is that the GPU is getting better precision; because it might do blending in f32. But I don't know.

[DJMcNab]

The messages for those test failures could be a lot more helpful - linking to the diff image, for example. We probably should also set up #652 for this (and update it to reflect #892 as well)

[LaurenzV]

Using my branch I'm only getting 14 failures with 0 tolerance, and they are mostly at the edges, e.g.

However, there is one image where the difference is quite big:

Will see whether I can investigate what is going on here.

[LaurenzV]

Guess we just got unlucky... At some point we end up with the number 0.7509804 on the CPU, which multiplies to 191.500002, and therefore gets rounded up to 192. On the GPU, it gets rounded down apparently. Not sure there is much point in investigating this further though, so I think the best thing would be to just stick with the default tolerance threshold of 1.

It might still be nice to merge the change in this PR at some point, for the sake of consistency across the renderers.