bug(common/models): predictive text suggestion volatility

>Something interesting at 0:13s, 1:15s and 1:21s.
>
>https://github.com/keymanapp/keyman/assets/28331388/349538ad-de74-4eef-99e2-e3172e7e8b89
>
_Originally posted by @MakaraSok in https://github.com/keymanapp/keyman/issues/8544#issuecomment-1545392577_

Upon first glance, I think the "interesting points" seem to be those where the user has typed most rapidly. I think this is related - less time between keystrokes means less time to predict... and so the predictive-text engine's result quality drops due to having less resources.

Note: in the first and third points, the banner simply goes blank. In essence, no good suggestion could be found in time. Granted, `tooo` should be pretty dang close to `too`, so in theory, this still shouldn't be a common occurrence. The second one is... weirder, though it is worth noting that it is also just one keystroke off: **Y**ooo vs too~o~. It seems odd that the former would be found first by the prediction-search, though not impossible.

Either way, it may be worth some investigation to confirm these suspicions and to see if there's a reasonable way to address these concerns.