Adds a thread-local interner of layouts to LayoutCache, and updates all
references appropriately.
This is a bit suboptimal for single-threaded workloads that will look at
creating layout caches again, like the REPL, but I think that's okay for
now - since the global interner will be uncontested for those workloads, it
should still be plenty fast to access the interner, even behind a lock.