Track benchmark performance in CI with CodSpeed #2587

sqlparser_bench never runs in CI, only clippy checks it. The benchmarks once measured error paths without anyone noticing (#1559), and the six parse_*_chain groups that guard the exponential parse-time fixes (#2344, #2350, #2352, #2372, #2392) would not catch a regression either.

The proposal is to run them with CodSpeed in simulation mode, which counts instructions under Valgrind on ubuntu-latest, so results stay stable on shared runners.

Two ASF prerequisites.

• CodSpeedHQ/action back on the allow-list (removed as unused in Remove 14 unused wildcard actions from allowlist (part of #666) infrastructure-actions#1036).
• The CodSpeed GitHub App installed on this repo through an INFRA ticket, as OpenDAL did.

apache/arrow-rs#9975 and apache/datafusion#21165 plan the same setup. OpenDAL dropped CodSpeed in apache/opendal#5626 over false positives from I/O-bound benchmarks, while parsing is CPU-bound.

• Get the action approved allowlist: re-add CodSpeedHQ/action v5.2.1 infrastructure-actions#1327
• Swap criterion for codspeed-criterion-compat in sqlparser_bench/Cargo.toml. Bench sources and local cargo bench stay unchanged.
• One workflow on push to main plus a nightly schedule, and PR runs once the numbers prove stable.
• About 30 benchmarks, so one job.