Add caching to ConfigurableLanguageManager::getLanguages()

Seeing two calls to that that aren't statically cached. Once with STATE_CONFIGURABLE, very early, in \Drupal\language\EventSubscriber\LanguageRequestSubscriber::setLanguageOverrides() after/as part of language negotation. And the second one is the EntityConverter param upcasting (through EntityRepository => ContentEntitBase::getLanguages(), which explicitly uses STATE_ALL.

The second cache creates the same information, but mostly hits static caches (CachedStorage and ConfigFactory).

Started with a basic implementation, lets see if this causes any test fails. Also didn't yet add static caching for $languages and no DI either.

To get the tests to pass, this required a single manual cache clear in a test that's explicitly messing with the default language service in the container on the fly. I think that's fair.

This saves an early bootstrap query on every single multilingual site requests that isn't an internal page cache hit. When using redis, that means the amount of queries go down from 6 to 5 on a dynamic page cache hit:

What's strange is that \Drupal\Tests\demo_umami\FunctionalJavascript\OpenTelemetryAuthenticatedPerformanceTest::testFrontPageAuthenticatedWarmCache isn't failing, because that's exactly the query that should go away. Ah, I think those tests don't run by default. I also wasn't able to run it manually, go this error:

This job is stuck because of one of the following problems. There are no active runners online, no runners for the protected branch , or no runners that match all of the job's tags: performance-test-runner 

Locally, it 's failing, but not only because that query is gone, also because I'm getting a race condition on automated_cron. Reliably got this fail 3 times:

1) Drupal\Tests\demo_umami\FunctionalJavascript\OpenTelemetryAuthenticatedPerformanceTest::testFrontPageAuthenticatedWarmCache
Failed asserting that two arrays are identical.
--- Expected
+++ Actual
@@ @@
     0 => 'SELECT "session" FROM "sessions" WHERE "sid" = "SESSION_ID" LIMIT 0, 1',
     1 => 'SELECT * FROM "users_field_data" "u" WHERE "u"."uid" = "10" AND "u"."default_langcode" = 1',
     2 => 'SELECT "roles_target_id" FROM "user__roles" WHERE "entity_id" = "10"',
-    3 => 'SELECT "config"."name" AS "name" FROM "config" "config" WHERE ("collection" = "") AND ("name" LIKE "language.entity.%" ESCAPE '\\') ORDER BY "collection" ASC, "name" ASC',
+    3 => 'INSERT INTO "semaphore" ("name", "value", "expire") VALUES ("cron", "LOCK_ID", "EXPIRE")',
+    4 => 'SELECT "expire", "value" FROM "semaphore" WHERE "name" = "cron"',
+    5 => 'INSERT INTO "watchdog" ("uid", "type", "message", "variables", "severity", "link", "location", "referer", "hostname", "timestamp") VALUES (0, "cron", "Attempting to re-run cron while it is already running.", "WATCHDOG_DATA", 4, "", "LOCATION", "REFERER", "CLIENT_IP", "TIMESTAMP")',
 ]

I think the non-cron-on-installer change really did a number on this test. Others maybe too, didn't try to run them. The umami based test is special because it's first cron run is really slow. There's a lot to index there.

I tried a number of things such as the wait for terminate event (adds another query for the lock and wasn't reliable enough either), explicit lock wait(), sleep(). The one thing that seems to reliably help is explicitly running cron before the first web request. Maybe the performance tests should just run without that module?

One question is whether we can avoid the cache tag. We have a variable amount of cache keys, but the amount of variations is limited, we might be able to do a bulk delete this. Not sure yet.

Based on discussion in slack, I've simplified the caching to just cache the listAll(). Then it's a single cache, we can delete that in reset(), no more test changes necessary, no extra cache tags.

Also found #2686079: isMultilingual() loads all languages just to check if there is more than one again, very similar issue, proposed there a different fix, I think this is simpler and good enough.

Rebased, added a comment and updated the new OpenTelemetryNodePagePerformanceTest assertions. Note that it adds two cache gets because the language list gets initialized twice with different flags. I think that's fine for the fast chained bin and doesn't require static caching for this.

Rebased.

Rebased and squashed the commits together to make future rebases easier.

Replied.