Reduce concurrent adM requests

[pjenvey]

When first populating or refreshing contile's cache with content from adM, multiple requests for the same content can be triggered simultaneously. Many of these request are redundant, which:

1. add extra work for contile to process and write them to the cache (see Avoid the cache "dog-piling" effect #46 for potential cache side effects)
2. cause extra traffic to adM (particularly problematic during downtime conditions Rate limit calls to adM in downtime conditions #220)

With Contile now handling image hosting its processing task is potentially lengthier, which could result in even larger number of redundant concurrent requests triggered (since the cache doesn't fill as quickly).

To improve this: when an adM request is triggered Contile could write this event to its cache, letting subsequent requests act more intelligently. Instead of spawning more requests they could instead: wait for that adM request to finish (by simply 503ing the client or actually waiting over an spmc channel for its completion) or by simply returning stale content until the request finishes.

┆Issue is synchronized with this Jira Task