Since Go doesn't support unions, and because the translation format is a bit
wacky with the same field having multiple types, we must resort to
introspection to switch between single-item translation (for singular), and
multi-items ones (for plurals).
Previously, introspection was done at runtime, which is not only slow, but will
also only catch typing errors while trying to use the translations. The current
approach is to use a struct with a different field per possible type, and
implement a custom unmarshaller to dispatch the translations to the right one.
This should marginally reduce the memory consumption since interface-boxing
doesn't happen anymore, speed up the translations matching, and enforce proper
typing earlier. This also allows us to remove a bunch of now-useless tests.
Instead of having a switch-case returning a function to be executed, it's
simpler/faster to have a single function containing a switch-case. It also
allows to group languages with identical plural form in a single
implementation, and remove the "default" guard value, as switch-case already
have a `default:` case.
While doing some profiling for #2900, I noticed that
`miniflux.app/v2/internal/locale.LoadCatalogMessages` is responsible for more
than 10% of the consumed memory. As most miniflux instances won't have enough
diverse users to use all the available translations at the same time, it
makes sense to load them on demand.
The overhead is a single function call and a check in a map, per call to
translation-related functions.