Benchmarks: fix test_export_client_twin after #456, and run the benchmarks/experiments tests in CI - #474

Follow-up to #456, which merged with test_export_client_twin.py::test_the_setup_and_the_sparse_table_say_what_the_rows_stamp failing. Nothing in CI ran benchmarks/experiments/test_*.py (checked: no workflow references them; only lint-workflows.yml mentions the directory), so nobody saw it.

• Fix: Benchmarks: the served ArcadeDB arms leave -Xms unset, like the embedded arms #456 made the exporter's jvm_heap sentence say "ArcadeDB, embedded and served, sets only the maximum heap"; the test still asserted the old "ArcadeDB embedded sets only the maximum heap". The assertion now matches the sentence.
• CI: new test-benchmarks-experiments.yml: standard ubuntu-24.04 runner, only for changes under benchmarks/** (and the workflow itself), concurrency group with cancel-in-progress on pull requests, uv installed by the same pinned script as the other workflows, then uv run --no-project --with pytest --with numpy --with pandas --with pyarrow --with requests python -m pytest -q -p no:cacheprovider . from benchmarks/experiments. 10 minute timeout; the suite takes about 75 s locally. No Docker build, no timed cell. Registered in sync-upstream.sh's fork-only allowlist like docs-build.yml.
• Result on this branch locally: 971 passed, 24 skipped, 0 failed (on main before the fix: 1 failed).

🤖 Generated with Claude Code