change(web): improve optimization of prediction-generation

[jahorton]

As noted on #16450, the "correction" portion of suggestion generation is susceptible to the host device's performance capabilities. Keeping the process well-optimized helps ensure that reasonably-expected corrections actually surface.

At present, correction and prediction operate as follows:

1. A correction-search process searches for the best-fitting corrections for the existing context. The "optimization condition" is built from two parts:
   1. The combined probability of the keystrokes selected at each stage, as taken in a vacuum. These probabilities are based upon the distance from the center of the key to the actual touch-point of the user's finger at the time of the key event.
   2. A "probability" based directly upon the number of "edits" applied to the context in addition to the user's keystrokes - an "edit distance", if you will. Potential edits:
      1. Insertion of an extra character that was never entered
      2. Deletion of a character that was mistakenly entered
      3. Wholesale replacement of an entered character
      4. Reordering two keystrokes so that their text-edits are applied in inverse order
2. Once viable corrections are found, the engine then generates predictions based upon them.
   • Note that the frequency of predicted words, and thus their "probability", is not considered for the search.

Given that proper handling of transpositions must already overcome a notable penalty due to edit-distance cost, we should probably include an aspect of "minimum prediction cost" within the correction-search optimization condition. This way, transpositions toward high-frequency words gain stronger preference than edits toward lower-frequency words, helping to ensure that the highest-priority transposition corrections are able to surface more reliably.

This would result in an optimization condition built from three parts:

1. The combined probability of the keystrokes selected at each stage, as taken in a vacuum.
2. A "probability" based directly upon the number of "edits" applied to the context in addition to the user's keystrokes - the "edit distance".
3. The lexical probability of the most likely word prefixed by the corrected version of the text.

[jahorton]

See also #16534.