Update the CodSpeed benchmark harnesses and runner (#30323)
## Summary
We see cases where the codspeed result differs based on the order in
which the benchmarks run, because they all run in the same process (and
use the same allocator). Newer codspeed versions run each benchmark in a
separate process, protecting them from "warm allocator" effects.
This PR updates `cargo-codspeed` to 6.0.0 and the CodSpeed action to
5.6.3. Each simulated Criterion or Divan benchmark now starts in a fresh
process, so allocations in one fixture can no longer change the
instruction count of the next.
## Test Plan
Verified that the linter fixture and all four module-resolution inputs
each ran in a separate process, and that the updated dependencies meet
the seven-day cooldown.
