Bucket sampling for preference experiments

[Osmose]

This is related to #584 in that we require per-action server logic. In this case, we need to make the server compute a custom filter expression for recipes that are a certain type of action, and append it to the other filters on the recipe already.

preference-experiment actions will use this to filter by sampling buckets. Users will choose the number of buckets they want, and the server will compute a filter expression that:

• Splits the userbase into 100,000 buckets, and reserve even chunks of that space for each branch.
• Uses the user's unique ID and the experiment slug for the hash
• At each of those start points, matches users for X buckets, where X is defined in the recipe's arguments per-branch.

[Osmose]

Our approach for this has changed since the bug was initially filed, after talking with Sunnah and Gregg. Now, preference experiments specify a count of how many buckets out of 100,000 that they want to sample, and branches specify the ratio of matched users that they should match. Thus, the filtering for this simplifies to doing a normal, independent bucket sample of users based on that count.

We can probably accomplish this within the preference experiment action too, which removes the need to refactor the backend to support this for all actions. Yay!

[Osmose]

Per the discussion in #650, this is being moved to a non-blocker since we can write a filter expression that accomplishes this easily and has the same effect as the PR I put up.

We still should address what UI we want to making this kind of filtering accessible to the audience we eventually expect will be using Normandy to launch preference experiments, but for an MVP it's not necessary.

[Osmose]

I think the solution to this lies in #316 and related efforts to simplify sampling controls. This issue seems redundant now.