[Bug]: Beats setup tasks (filebeat/metricbeat) inconsistent with auditbeat

The three Beats run their setup task (loads index template, ILM, ingest pipelines, dashboards into Elasticsearch) inconsistently, and each has gaps. Auditbeat is the closest to correct and should be the reference for the other two.

Problems

• metricbeat uses bare metricbeat setup (= index-management + dashboards + ML, but no pipelines), with a marker misleadingly named pipelines_created.
• filebeat is not gated on the Elasticsearch output (fails on the default Logstash output); passes --pipelines --dashboards but no --index-management; modules live in modules.d/, so setup --pipelines finds none (needs --modules); marker is per-module.
• auditbeat uses a run-once-forever marker, so modules/filesets added later never get their pipelines loaded.

Target

• Gate all three on beats_<beat>_output == "elasticsearch" and run_once.
• Always use explicit flags, never bare setup, so the task is self-documenting: setup --index-management --pipelines, plus --dashboards only when Kibana is configured. --index-management is safe (setup.template.overwrite: false).
• Rename metricbeat's pipelines_created marker to an accurate name (e.g. setup_done).
• Replace the run-once-forever marker with re-run on change: register the module/config task and run setup when: <register> is changed.
• filebeat: add --modules (or adopt the inline data-driven design) so module pipelines actually load.
• Drop "(experimental)" from the beats_filebeat_modules description.

Background

With the Elasticsearch output the index template and pipelines auto-load on first run anyway; with the Logstash output nothing auto-loads, so setup is only really needed there — yet the tasks gate on the ES output and skip exactly that case.

The module handling is redesigned data-driven in #510 / #505 / #516 (inline, one list = desired state), which simplifies this further (inline → setup --pipelines sees the modules directly, no --modules). This issue is the fix for the current architecture.

Start with the Beats setup tasks and the module/config tasks referenced in the issue, then review the data-driven redesign in #510, #505, and #516. Compare filebeat, metricbeat, and auditbeat against the explicit setup flags, output and run_once conditions. Done means all three setup flows use the requested gating, markers, rerun behavior, module handling, and description change.