Tokio shouldn't be used for local reads

[myrrc]

Starting in #9175, we need to remove tokio from local reads as they don't scale well and burn CPU significantly on 128+ cores.

[AdamGS]

I don't think that's right. The root cause here is that because the way our IO works we can't do multiple reads in a single pass, which forces us into a suboptimal IO pattern on any fast storage.

[myrrc]

Yeah, good solution is vectorized reads, and I have a prototype for that, but making it production-ready would take too much capacity so it's lower priority.

[AdamGS]

If you're not going to push it through, I can take this issue up I just need to understand where it stands.