better explain how scores are computed / which columns can be compared

The scores of different benchmarks are computed in different ways (e.g. distributional match for Marques2020, r^2 for neural predictivity, r for Rajalingham2018-i2n). This is not immediately clear from the website, which can make users compare scores across columns without being aware of the differences.

(from @AlinaPeter, @kohitij-kar, @ratanmurty)

Start by reviewing the website areas that display benchmark scores and identify where score computation and comparability are currently described. Document the stated methods, including distributional match, r^2, and r, and clarify which columns should not be compared. Done means users can distinguish the metrics before comparing scores.

Written by the indexing model from the issue text.