feat(state,parse): Perform partial updates on StopEvent

[runkelcorey]

Asana Ticket: 🐞 State.StopEvent times out on refresh

What changes does this PR propose?

Changes refresh strategy for StopEvent from a full load to an incremental update by:

1. Implementing handle_new_state callback that filters incoming stop event JSONs for records that are newer than a configurable timestamp (defaulting to 2 hours, the same that Flashback aims for)
2. Implementing post_commit_hook callback that evicts records from the mnesia table if they are older than the configurable timestamp

Since StopEvent is a :set table, incoming values with the same keys will overwrite existing values.

How were these changes validated?

1. Added tests to State.StopEvent and Parse.StopEvent that different cases of partial updates
2. Deployed to dev-blue and saw dramatic differences between the refresh times in dev-blue and dev-green (which is running the current, full-load strategy):

measure	dev-green	dev-blue	𝚫
5th percentile	15,681ms	11ms	-99.9%
median	17,459ms	14ms	-99.9%
95th percentile	19,957ms	27ms	-99.9%

These are crazy improvements so I feel like I'm missing a logging statement to capture the full overhead of the partial load. However, we should expect something like this: of the 116,740 events in the stop events dataset I downloaded yesterday, only 356 (0.3%) were updated or inserted the last 30 seconds of the dataset.

As a sanity check, I also counted the number of times each of these function were executed: since this is an event loop dictated by the time it takes to refresh data + sleep time, the more times something is invoked can indicate performance impacts. dev-blue completed 1,059 refreshes of StopEvent in 30 minutes, which translates to once per 1.7s, while dev-green completed 432 refreshes in the same time period, a rate of one refresh every 4s.

Performance-wise, these changes seem very impactful.

What questions should reviewers consider?

1. This logic duplicates how we create the dataset in Lamp: upsert new records and evict old ones; is this a sign that the API is doing too much?
2. Do you think that the logic for a full load is really just a special case of a partial load?
3. Stylistically, should timestamp be a datetime or is it ok to have as a Unix timestamp?
4. Is there an better way to evict that would work better here? Like, to fully embrace the actor model, I imagine having each key track its own eviction but not sure if Mnesia invalidates some of those actor model assumptions