IO/waiting for forked jobs can be done in background threads since the lock is released on some select function calls.
So while incremental GC certainly does bring down latency caused by the GC it does not bring down latency for tasks that could be parallelized.
The fact that those threads do not run in parallel is an entirely unrelated issue. Parallelism is not an essential quality of multi-threading.
Here's my basic understanding:
MRI is a single process, with multiple threads... They can't run in parallel because of the global interpreter lock (which is used because the code in MRI is not thread safe? Meaning each thread can touch the same memory and may leave it in a broken state for another thread, so we lock access to it first...).
So: there are multiple threads, I am guessing each with their own task (maybe one to run the code, one to do garbage collection, ... I have no idea), but because of the GIL they can't be parallelized. Still multi-threaded anyway.
What makes threads interesting is parallelism; if the goal is merely cooperative or pre-emptive multitasking, where there's no concurrency, then you're "effectively" not multi-threaded as per any modern expectation.
The answer is that it's often useful to manage multiple things at the same time, even if we can't or don't want to actually do them at the same time.
We may want to use blocking IO for the simplicity it gives us compared to asynchronous IO, but allow other parts of the system to continue to run during that blocking.
We may want to let threads and the kernel manage our context and context switches for us, instead of manually storing and restoring state, because it can be simpler that way.
They largely weren't. POSIX threads weren't standardized and widely adopted until the 90s, and it wasn't until multi-processor systems became more prevalent that threads saw wide adoption at all.