No translation round-trip on every page load

Commit message

[PERF] base: skip fetching up-to-date translations

web, http_routing, website

On every page load, in the backend and on the website, the localization
service fetched the translations (/web/webclient/translations or
/website/translations) to check whether the ones stored in IndexedDB
were still current, even when they were. That check cost a full request
per page view, about 6-7 ms of worker time.

The session info now contains the hash of the translations that the
translations endpoint would return, so the client only fetches them when
its stored hash differs.

The hash is only taken from the ormcache, with the new
ormcache.get_value(), which returns a cached value without computing it:
building the hash requires building all the translations, which must not
slow down page rendering. On a cold cache the session info has an empty
hash, so the client fetches the translations as before, which fills the
cache for the next page loads. Website pages served from the page cache
get the current hash, as the page may have been cached while the hash
wasn't.

The default modules and the language check (a language that isn't
installed becomes None) used by the translations endpoint are moved to
ir.http._get_web_translations_modules and _get_web_translations_lang,
next to _get_web_translations_hash, so the endpoint and the session info
always hash the same modules and language.

With 4 workers, 5 languages and the translations fetched in the
background, the worker time per page view goes from 20.5 to 14.1 ms on
the website and from 21.9 to 13.6 ms in the backend, without changing
the cost of the page itself.