When we're searching for a color that has a contrast 1.07 (for example) against the seed, we add a "topup" constant before the search, so that after all rounding and approximations the contrast is strictly higher than 1.07. Otherwise the result contrast could be quite lower:

• we're searching for contrast with tolerance 0.004 (CONTRAST_EPSILON), so the found color could well have contrast just 1.066.
• the resulting color is serialized as a hex value, which has 1/255 step. Contrast between adjacent steps is 1.008 to 1.016. It's nonlinear, and highest in the middle of the range, around #4a4a4a. So, when the color is rounded down, the real contrast can be as low as 1.058, a 0.012 difference.

The topup constant is here to prevent this, but until now it's been way too high. This PR changes the 0.05 value to the theoretically best 0.012 value.

This changes the default ramps quite visibly. For example, the surface3 background color, used for the surface-neutral-strong token, has changed from #fff (pure white) to #fcfcfc. That's because it's supposed to have contrast 1.02 against the seed color #f8f8f8. But after the 0.05 topup, we were aiming for 1.07, a much lighter color. Now we aim for just 1.032, which is satisfied by #fcfcfc and we don't need to go all the way to pure white.

@jameskoster @poligilad How does this affect your recent efforts to adjust the seed color from #f8f8f8 to #fcfcfc? This PR shows that:

• attempts to fine tune the exact colors are somewhat futile, as the colors are generated by an algorithm, and the algorithm can change.
• the algorithm is not perfect, and maybe we want algorithm changes instead of specific color fine-tuning.