fix: announce only what concurrent override writes stored

[JDCyllene]

What this pull request does

Point 4 quater of the changes the Sylius integration asked for, engine side only.

OverrideWriter::remove() read the override, then deleted it, without a lock. Two removals of the same key at once (two admins resetting the same entry) both found the row: the second DELETE matched nothing, but both invalidated the cache and dispatched OverrideRemovedEvent, so the host purged and notified twice for one removal.

The row is now read locked (TranslationOverrideRepository::lockOneByKey(), PESSIMISTIC_WRITE) in a transaction of the writer, and the effects follow its commit (withEffectsAfter(), as save() does): the second removal waits for the first, finds nothing, and announces nothing.

After review, save() is fixed the same way: it read without a lock, so a save meeting a removal of the same key found the row, its UPDATE waited and matched nothing (Doctrine does not check how many rows a write touched), and OverrideSavedEvent announced a value that was not stored. A row the save finds is now read again with lockOneByKey(): the save waits for the removal and stores its value in a new row. An absent key is never read locked: on MySQL and MariaDB a locked read of an absent key takes a gap lock, and two first writes of different keys would wait for each other, or deadlock. The race of ADR 0016 (two first writes of the same key) is unchanged. The batch writes (writeMany(): import, batch approval, removeMany()) keep reading without a lock, a limit documented in docs/concepts/manual-override.md.

removeOverride() keeps its signature (void): telling the caller whether a row was removed would change an interface a host may implement, so it is left for later. docs/concepts/overrides.md, docs/concepts/manual-override.md and CLAUDE.md are updated.

Points to watch

• Tests pass (vendor/bin/phpunit): SQLite, MariaDB 11.8, PostgreSQL 17, without intl and with the lowest dependencies
• PHPStan and PHP-CS-Fixer are green (vendor/bin/phpstan analyse, vendor/bin/php-cs-fixer check)
• UPGRADE.md is up to date if the public contract changes: the contract does not change
• A decision taken or reversed has its record in docs/adr/: none, this applies 0013 (stale entities never decide)
• No real API key appears in the code, the tests or this description

No change to the entity mapping nor to the configuration tree: the Flex recipe stays as it is.

For review:

• ConcurrentOverrideWritesTest runs a second PHP process (proc_open) that writes in a transaction it keeps open for one second while this process writes. Three cases: two removals of the same key (fails on the previous code with MariaDB and PostgreSQL: two events), a save during a removal (fails on the previous code: nothing stored), and two first writes of different keys, both through the bundle's save() (fails on MariaDB when the absent key is read locked: the second waits one second). Each case checks how long the write took, so a slow runner fails the test instead of passing it without proving anything. Skipped on SQLite, which has no row lock: it runs in the databases job of the CI.
• A removal now opens a transaction of its own; called inside the host's transaction, it nests in it, as save() already did.