.github: Only test benchmarks, don't run them

[mxinden]

Instead of fully executing benchmarks, i.e. running multiple iterations,
each measured and recorded, only test that they compile and run them
with a single execution to make sure they work. The benefit is a reduced
CI runtime.

Flag documentation:

To test that the benchmarks run successfully without performing the
measurement or analysis (eg. in a CI setting), use cargo test --benches.

https://bheisler.github.io/criterion.rs/book/user_guide/command_line_options.html

The above assumes that (a) the benchmark results from CI are likely
noisy and thus (b) no one actually looks at the benchmark results.