Lock congestion when the same interned value is used in a hot function

[MichaReiser]

This is something that has come up in ty, see astral-sh/ty#968

If there's a hot function that calls new on very few interned values (let's say two!), then the threads start blocking each other because intern_id uses a Mutex sharded by value.

We could consider using a RwLock but that comes at the cost that the cold path (a new value or a value that hasn't been interned yet) becomes more expensive.