[Feature] Extending `dag_export`'s flexibility so that users can use it without a cost model

[Selvomega]

Current State and Why It Is Unsatisfying

--post-asap without --planner-cost-json runs nothing.
dag_export.rs:1500:

let results = if let Some(document) = planner_cost.as_ref() {
    let model = ExportPlannerCostModel { document };
    run_post_asap_with_progress(&lowered_queries, progress, &model, Some(&model), ...)
} else {
    eprintln!("dag_export: --post-asap requires complete deployment-owned \
               physical-plan evidence; exporting the raw plan only");
    raw_only_post_asap_results()          // three empty Vecs
};

The gate exists because a post-ASAP export asserts a cost-ranked decision, and
dag_export owns no deployment evidence with which to defend one.

However, a user with no deployment yet still wants to see what ASAPPlanner does
with a workload — which replacements it finds, which it commits to, what the
DAG looks like — and that needs the structure to be real, not the numbers.

Proposal

1. Behind --default-cost (mutually exclusive with --planner-cost-json), run
   the search and pass &DefaultCostModel to global_selection — the two lines
   5dbf4d9 deleted.
2. Keep export_model as None, so every cost is CostSource::Unavailable
   with value: None. workload_cost_summary already degrades to an
   unavailable summary; the viewer renders Not estimated.
3. Absent both flags, behavior is unchanged.
4. Update the stderr line and tools/dag-viewer/README.md, which currently
   states that without the document --post-asap exports the raw graph only.

Why This Is Not What 5dbf4d9 Removed

That commit deleted two things:

Removed	Restored?
winner_cost_annotations emitting CostAnnotation::modeled(...) from AnalyticalCostModel	No
analytical.map_or(&default_model, ...) choosing which model ranks	Yes

DefaultCostModel's number was never an exported cost. The old
winner_cost_annotations had two outcomes: calibrated CostUnits, or
Unavailable — "structural node counts are intentionally never used as a
fallback." The commit's real work was making the evidence document complete and
fail closed. Untouched here. The usual objection — that dag_node_count plus an
"illustrative, not measured" weight table is indistinguishable from a real
cost — does not apply to a number that is never serialized.