Update CompletionSuggestion bucket selection
The data collected by this experiment doesn't make sense. On review
it was suggested to make a few changes:
- don't use Math.random(), instead sample from the random pageId var.
- unrelated to unrelable data collection, but removed logId as not necessary.
- Ensure pageId is only generated once, and not multiple times, when participateInTest is called multiple times on a slow connection.
- no longer blocks initial suggestions waiting for mw.loader.using