Saying that the time to switch is small is probably not the most important part of the story, because any independent program threads which can share resources have to deal with contention issues - where two things seek to use the same address simultaneously. This is a problem first formulated by
Edsger Djikstra. In some cases the modern implementations of solutions can actually take a long time to clear (I've seen "milliseconds" being referenced as the amount of time it takes a spinlock to clear through, for example). As a guess, the main game execution thread of MAME benefits on the whole from offloading some rendering and polling tasks to another core - but as
this thread explains this idea hasn't been well executed so far (in fact Haze suggests not using multithreading with MAME at all!) As usual, the main issue is not so much speed but rather safety and troubleshooting - and of course in this case possibly strange timing issues as well. Back to basics, I guess - people have pushed Amdahl's Law and multiprocessing in search of more performance and time savings, but Djikstra (the same as in his paper "Go To Considered Harmful") was more interested in safety.
As a matter of semantics, the Saturn has never been called a multi-core system because cores refer to functional units which require additional circuitry to work singly. Each CPU on the Saturn has its own bus interface, but each core within a multi-core CPU shares parts of that main interface. But there doesn't seem to be any standard definition of this, just a bunch of conventions that carved out their own niches (MP, SMP, etc.) before people started talking about multi-core.