Save results as JSON and run YJIT and ZJIT variants

[JuanVqz]

Summary

• Benchmark results only existed as text in job logs and as ips.fastruby.io reports, which don't record the Ruby, OS or machine. Each result is now also saved as JSON, together with where it ran, so a future results site can show only comparable numbers side by side.
• Every CRuby run was interpreter-only. YJIT is built into Ruby 3.2+ but off by default, while many production apps run it (Rails 7.2 turns it on for new apps on Ruby 3.3+). CI now also runs YJIT and ZJIT variants.

Changes

Step 1: results as JSON

• docker/collect_results.rb, loaded through RUBYOPT when RESULTS_DIR is set, so benchmark files stay untouched. Each Benchmark.ips report goes to RESULTS_DIR/<label>/<benchmark>.json with the Ruby, the variant and flags, the PR's own commit and number, and the environment: arch, CPU, core count, OS, glibc, compiler and optflags. A field that can't be read is null, and the others are kept.
• CI sets RESULTS_DIR on every benchmark job and uploads a results-<label> artifact.
• Works on every Ruby in the matrix, back to 2.1.

Step 2: YJIT and ZJIT variants

• 8 more CI jobs: +yjit on ruby_3.1 to ruby_head, and +zjit on ruby_4.0 and ruby_head. Each is its own job (variant: [''] plus include, so the entries don't merge into the plain jobs).
• The flags go through RUBYOPT, JRUBY_OPTS or TRUFFLERUBYOPT, so every benchmark process the Rakefile starts inherits them. The run stops if the variant isn't on, instead of quietly benchmarking the interpreter.
• Only plain Rubies share to ips.fastruby.io, because shared reports don't record the Ruby or its flags.
• CONTRIBUTING shows how to run a variant and keep results locally.

Test plan

• Every flag reaches child processes through its variable: YJIT and ZJIT (RUBY_DESCRIPTION), plus JRuby --dev and TruffleRuby with compilation off, which are supported but not in the matrix.
• Results written on ruby_3.4+yjit, ruby_4.0+zjit, ruby_2.1, jruby_9.1, jruby_head, truffleruby_22 and truffleruby_head, through both file mode and the full Rakefile run. ruby_3.4+zjit stops with "invalid option --zjit" and writes nothing.
• The amd64 ruby:3.1 image has YJIT (+YJIT [x86_64-linux]), so ruby_3.1+yjit passes its check on CI.
• Review fixes: arm64 records CPU implementer and part codes instead of null; a missing /etc/os-release or ldd leaves only those fields null; an empty RESULTS_LABEL falls back to the engine and version, plus the variant (ruby-3.4.10+yjit), so local plain and YJIT runs don't overwrite each other; RUBY_VARIANT without flags stops with "Variant yjit is not on"; a PR run records the PR's head commit and number instead of GitHub's temporary merge commit.
• Matrix expansion checked against GitHub's documented include rules: 26 jobs, the 18 plain ones unchanged. actionlint is clean.
• This PR changes the workflow and docker/, so its CI runs the full suite on all 26 jobs. It is also the first real check of the artifacts and of how many CPU models one run gets.