stdlib #[bench]s are (possibly) never run

[huonw]

It seems to me that the only way for #15833 to be an issue is if we never run #[bench]s anywhere during a PR merge (or on travis etc.).

Maybe the test harness could run #[bench]s as normal tests during a normal run (i.e. execute the code exactly once, not bothering to time it).

[alexcrichton]

I believe they are all currently compiled, but never run unless you pass --bench which is quite rare.

Running a benchmark for one iteration seems quite reasonable!

[steveklabnik]

Triage: don't think anything has changed here.