fix: rewrite text scanning

[birtles]

Completely overhaul text lookup to cover more cases and be (hopefully) easier to
extend in future (so we can tackle #1130).

Fixes #990.

As mentioned in the discussion for #990, this has a negative performance impact
for most content. This is due to the extra distance calculations performed on
each lookup which is needed to avoid returning a cached position when the
distance has substantially changed.

For some content such as asahi.com, however, there is a net performance
improvement due to the way we handle repeated lookups.

For now this is probably acceptable, but if not, we will need to find a way to
cache the point used as input for the previous lookup and avoid doing distance
calculations if we determine we are close.

[birtles]

This probably needs testing with Safari before merging

[birtles]

Unit tests are failing on Chrome on CI, presumably because we no longer apply the mid-character adjustments to synthesized input text nodes there. That's probably easy enough to add back.

[birtles]

Performance impact is now fixed (in the majority of cases this is faster). Just need to test on Safari which means I need to wait until I can get near a Mac.