We do this by reducing precision proportionately to the value, rather than simply rounding to a certain number of decimals or similar, so that small sizes maintain finer precision. This is particularly important for cases where content specifies tiny font-size values with the expectation that a large scale() factor will later be applied. A perftest with lots of "random" floating-point font sizes is added, based on the testcase in the report. (A few existing tests get minor adjustments to allow for the modified precision.) Differential Revision: https://phabricator.services.mozilla.com/D306131