ci(experimental): a merge wave queues one 2-hour GraalPy run per main push #1903

Every push to main starts its own Experimental Tests run, and nothing cancels the older ones. experimental.yml:31-33 groups pushes by run_id. Each run's Python graalpy-25.4 (experimental) job takes about 112 minutes.

On 2026-10-17 the owner merged 14 PRs within minutes. That left about 12 GraalPy jobs holding runners for two hours. Queued runs reached 65, and the open PRs' CI waited behind them.

Options:

1. Cancel superseded runs on main in experimental.yml only: group pushes by github.ref with cancel-in-progress: true. Experimental legs gate nothing (ci(python): run round-trip sweeps on latest only, move experimental legs to their own workflow (#1732) #1742), so the newest commit's run is enough. unit-tests.yml keeps one run per commit, because the unittest-ordering leg runs only on main.
2. Leave both workflows as they are. A merge wave costs about two hours of queue.

I recommend option 1. Which do you prefer?