Specify WHITE_COMPONENTS to machine precision

[tomcur]

These two are off by 1 ULP.

[waywardmonkeys]

Should we tighten the error bound in the white_components test?

[tomcur]

We could tighten it a bit, but we cannot detect 1 ULP errors. Depending on the exact computations that are performed, the largest magnitude of values during computation can determine the largest absolute error, which could in principle propagate to all components. The test itself only knows about input and output magnitudes, and not the magnitudes during computation.

The roundtrip test has a tight bound of 16 epsilons relative to the largest input magnitude, but cheats a bit by doing error checking in the linear sRGB space, and the bound may still be too tight if color space transformations do very fancy things. (OTOH, maybe if conversions can't reach that bound, the computations should be done in f64...)