Very interesting outcome.
Repository for the result and test source code: GitHub - newDINO/rust-thread-pool-compare: Some tests on common rust thread pool libraries · GitHub
It turns out chili, used with par-core and par-iter may be the best solution for parallel computation.
so far i have only checked the slice_add_assign source but i believe your benchmark is flawed you should use blackbox otherwise the optimizer will use the fact that the input is constant to avoid doing the work. also i think it would be more interesting to benchmark a scenario with an higher workload per item to have results that better reflect a real scenario