Remove continue-on-error for benchmarks in CI

[eregon]

• Otherwise new failures do not get noticed by the PR breaking it.
• Instead, just exclude the benchmarks known to fail.
• Relates to 6d43a4c. There has been some failures which were due to the harness assuming Ractor is defined, this CI job helps catch such issues. There is also CRuby-specific code in run_benchmarks.rb, so this tests it's properly guarded.

I'd like to remove the continue-on-error for benchmarks on TruffleRuby in CI, otherwise any regression there will most likely be missed. A concrete example is #375 didn't notice it broke all benchmarks on TruffleRuby due to the continue-on-error. I fixed that #442.

If it fails again, and it's not an issue of the code in this repository then please @-mention me (@eregon) and:

• If it's a single/few benchmarks then please add it to the --excludes list.
• If it's many benchmarks or transient, please re-add continue-on-error: ${{ matrix.ruby == 'truffleruby' }} in the name: Test run_benchmarks.rb step.

Is that reasonable, WDYT?

These benchmarks run with ruby-version: truffleruby, so the latest TruffleRuby 33 release, so that's more stable than e.g. using truffleruby-head. That also contains an important fix for Bundler install, which prevented to make this PR before.

Independently of this I will run the same benchmarks in truffleruby/truffleruby's CI to make sure we don't regress there on master, and that way new TruffleRuby releases shouldn't break ruby-bench's CI.