Give every benchmark the same shape and lint it

[JuanVqz]

Summary

• Every benchmark now has the same shape: default timing, a comparison at the end, and a Benchmark.ips that actually runs. A new lint enforces it in CI.
• Two files didn't fit. One measured with its own timing, and one never benchmarked anything.

Changes

Step 1: select-first-vs-detect.rb uses the default timing

• It was the only file measured for 20 s per report (Benchmark.ips(20)) instead of the default 5 s, since it was added in 87ff8f5 with no reason given. That made it about 4x slower than the others in every CI job.
• The result holds at the default timing, with margins no worse: detect is 3.82x faster than select.first on Ruby 3.4 (3.55x at 20 s), and 4.83x on Ruby 2.1 (4.53x at 20 s).

Step 2: remove concatenation_randomized.rb, which never ran

• Its three Benchmark.ips blocks sit inside methods that nothing calls, so running it only builds its random strings: no report, no comparison. CI logs show only its ruby -v line, and results collection (Save results as JSON and run YJIT and ZJIT variants #241) found nothing from it. The results in its comments came from a manual run.
• concatenation.rb already covers every method it compared (String#+, interpolation, String#concat, String#append). It came from Randomized string concat #216 for String concat contains dummy comparision.  #158 and was never linked from the README.

Step 3: lint every benchmark's shape in CI

• .github/scripts/lint-benchmarks.rb reads each file under code/ with Prism, without running it, and fails when a Benchmark.ips block:
  • doesn't call x.compare!;
  • sets its own timing (Benchmark.ips(20), x.time =, x.warmup =, or x.config with time/warmup in any hash syntax);
  • sits inside a method that never runs from the top of the file.
• A new lint job runs it on ruby_4.0. The benchmark jobs wait for it, so a failing lint doesn't hold 26 of the org's job slots. benchmarks-ok, the required check, fails when the lint fails.
• CONTRIBUTING's "Note on entry" lists the three rules.

Test plan

• The lint passes all 67 files, locally and under Docker ruby_4.0 (the command CI runs).
• 15 test cases (kept locally), checked against what each file really does by running it with a stand-in Benchmark.ips:
  • Flagged: each timing form, a benchmark only an uncalled helper reaches, a method named report that nothing calls, the uncalled one of two same-named class methods, a top-level call that reaches a different method with the same name, and files without Benchmark.ips.
  • Pass: Foo.new.run, class methods calling each other, class << self, A::B.run, inherited and mixed-in methods.
• benchmarks-ok passes only when changes and lint succeed and the benchmark jobs succeed or are skipped (all 6 combinations checked).
• actionlint and shellcheck are clean.
• This PR changes the workflow and .github/scripts/, so its own CI runs the full suite behind the new lint job.