Simplify backfill process

[duncanjbrown]

Depends on #20.

The old way 🪨

When bulk-importing existing records into BQ, we used to:

• get $batch_size records from the table using find_in_batches
• loop over them to build a heap of events in memory
• enqueue one SendEvents job with all the events in
• before the next batch, sleep for $sleep_time so we don't enqueue too many (??)

This wasn't great because we ran this via rake over ssh, and often the loop would outlive the ssh session timeout. Or the loop would crash for some reason with half the set enqueued. This meant we acquired some other "features", like being able to restart the process from a given ID (which would be helpfully logged when the process crashed).

The new way ✨

Break up the finding of the records with the building of the events

• get $batch_size records from the table using find_in_batches
• per batch, pluck all the IDs for the batch and give them as an array arg to a single LoadEntityBatch job

And in LoadEntityBatch:

• load the records for the IDs
• reduce them to a single payload with $batch_size events in
• send that payload

Also, cache the analytics data from the YAML so we don't read and parse it once per record

Advantages 📈

• atomic: either the batch succeeds or it doesn't, so we eliminate dupes
• much simpler outer loop so no need for all that recovery code
• sidekiq does all our retrying for us

Next 🔮

The outer loop should now be fast enough to run in-process from a web request, so we could try to build a little UI on top of it!