Presumably you'd still need some sort of synchronisation otherwise the memory will stay mutated just in one CPU's cache, you need to "flush" the updated memory so that other CPUs can then see it, but that's likely cheaper than executing the whole operation atomically... Basically, increments are fundamentally not atomic (i.e. composed of multiple separate operations) whereas reads (of a single word) are, so if all mutations happen from a single thread (assuming a sensible CPU architecture and cooperative compiler) there's nothing else that needs to be done to ensure atomicity, you only need to care about the "happens before" relationship.
But then I'm not a low-level programmer so the above is probably all wrong :)
A platform like the JVM disguises this because every read and write (even without 'volatile') use minimal synchronisation to avoid the worst aspects of the danger here. This ensures that programs that do it wrong (e.g. forget a 'volatile') are "only" incorrect, but not unsafe.
There's a variety of safe arena types that give allocation inside pools too: https://crates.io/search?q=Arena&sort=downloads