Avoid duplicate inserts when updating check runs

[timothysmith0609]

Refreshing a check run that is already stored attempts an insert before falling back to an update. This produces a duplicate-key SQL error for an ordinary update, even when the refresh succeeds.

Look up existing runs first and update them directly. Create missing runs inside a savepoint; if another writer creates the same run after the lookup, use a locking read to find it and apply the update. This preserves commit scoping, validations, callbacks, and the existing stale-webhook refresh behavior while keeping enclosing transactions usable after an insert conflict.

Adds regression coverage for updates without SQL errors, identical GitHub IDs on different commits, and a simulated concurrent insert inside a transaction.

Validation:

• The SQL-error regression fails against the previous implementation and passes with this change.
• Full SQLite test suite: 1,250 tests and 3,539 assertions passed.
• Check-run model tests on MySQL and PostgreSQL: 16 tests and 66 assertions passed on each database.
• RuboCop passes for both changed Ruby files.