Refactor grouping to share logic, track duplicates

[corps]

1. Track duplicates with a sentry metric for observability.
2. Generalize the neighbor search / insert logic by unifying two code paths. Effectively behavior remains the same as various subtle differences between the endpoints are accounted for in the new generalized logic.

[corps]

Duplicates will happen, and we shouldn't stress them too much. They are a natural consequence of CAP and can occur for numerous reasons:

1. Race conditions from retries
2. Serialization in job queues not as strong as one thought (in fact, celery cannot guarantee non duplicate jobs in the strong sense, only best effort it)
3. Future minor bugs

So observing but otherwise treating duplicates as expected system behavior simplifies behavior.