Improve AtlasDB crash and shutdown tolerance

[kantai]

One of the issues that #3047 surfaced relates to the resiliency of AtlasDB "attachments" events. These events are pushed from the chains-coordinator thread over a synchronous channel to the p2p thread, which inserts the attachments into the AtlasDB, and includes any newly-fulfilled attachments in the network result struct. This is pushed to the relayer thread, which notifies the event-listener.

When the node shuts down, these events can be lost. Instead, the coordinator thread should be emitting these events in a more resilient way (by writing them to the AtlasDB themselves).

[jcnelson]

The atlas DB writes have to happen in the same (logical) transaction as the transaction that writes out both the Stacks block to the chainstate and the events to the event observer.

At the same time, we need to determine the newly-arrived hashes, which get fed to the event observer, in the same transaction that feeds the block receipts to the event observer.

[kantai]

Closed with #3086