Improve speech performance, especially for lazy typesetting

[dpvc]

This is the PR that we discussed today in our meeting. I have reworked the handling of the speech-generation loop using the idea I mentioned toward the end: I record the return value from the setTimeout() command and use clearTimeout() to prevent the next iteration of the loop from running if the attachSpeech() render-action is called (e.g., the page is retypeset) while there are still MathItems waiting to be processed. That turns out to be much cleaner, and you can even eliminate the attachSpeechStart() method along the way. So I think these are good changes.

Just to have it in the record, this PR:

• Moves the Timing values to document options, where a page author can adjust them if needed.
• Changes the setTimeout() in the KeyExplorer to a direct call to attachSeech() if there is no speech already attached.
• Changes private to protected in case any of this needs to be subclassed.
• Adds speechTimeout to record the value of the setTimeout() calls for the speech loop.
• Uses that value to clear any pending loop calls if attachSpeech() is called while there are still MathItems in the awaitingSpeech array.
• Then directly starts the loop using setTimeout().
• Sets the process bit in attachSpeech(), indicating that the render-action has done its job (of starting the asynchronous loop). Without this, every typeset call will run attachSpeech(), even if it is not needed.
• The loop is made into a do-while loop so that at least one MathItem will be processed (avoiding a potential infinite loop if the threshold is set too small, like 0)
• If there are still more MathItems to process, we use setTimeout() to schedule the next loop and save the return value, otherwise, we clear the timer indicator and stop looping.
• The lazy handler is modified to override the MathItem's attachSpeech() method so that it is not performed until we actually typeset the math.
• We also increase the margin where the lazy loader will trigger typesetting so that we get a slightly larger head start on the speech processing.

[dpvc]

There is an issue with compiling the cjs version of the changes I made to the lazy handler. I will work out a fix.

[dpvc]

Ok, I made the fix.

[zorkow]

lgtm.