Function join2 

pub fn join2<A, B, RA, RB>(a: A, b: B) -> (RA, RB)
where
    A: FnOnce() -> RA + Send,
    B: FnOnce() -> RB + Send,
    RA: Send,
    RB: Send,

Two independent pieces of work.

The rayon arm forks; the spin arm runs them one after the other, because each half already spreads across the whole pool internally and nesting a region inside a region is the one thing CpuPool::run cannot parallelize. That is llama.cpp’s shape too: its threadpool runs one graph node at a time, full width.