perf(db): resolve packet endpoints at read time

[MrAlders0n]

#128 saved a full snapshot of every endpoint candidate (id, name, key, location) on each observation. In hindsight that was the wrong choice. Ambiguous one-byte hashes match ~7 nodes on average, and the same list got copied for every observer and every packet. On prod it grew to about 22 GB, roughly 60% of packet_observations.

This drops the column and resolves source/destination at read time, same as path hops already do.

• List pages (including scroll and WS reconnect backfill) resolve all their endpoints in one batch, two queries at most. Rows that used to show nothing now get endpoints too.
• Packet detail resolves every observation in the same single batch instead of per observation.
• Live WS events are unchanged.
• No API shape change.

Tradeoff: old packets show today's node names rather than the ones at capture time. With 3 to 7 day retention that barely matters.

Migration 038 drops resolved_endpoints. Space comes back as old rows age out, or right away with VACUUM FULL packet_observations.