sieval run --dry-run: report each task's per-stage resume plan #219

What

sieval run --dry-run (#217) reports each task's resume_plan outcome counts, but not the per-stage comparison that #211 introduced. A reader sees kept: 3, rereported: 3 without being told which stage changed, and has to derive whether a task infers from to_infer.

Proposal

Report, per task, the stage plan the runner decides from:

• stages: plan_resume's comparison of the recorded identities with this run's (same, differs or unknown per stage, identity_schema, legacy_grader), with rerun_from taken from the runner's own decision, so report names a re-report;
• needs_inference: whether the task sends any sample to its model under test, by the same test sieval run launches engines by.

A report-only change should read start: reuse, rereported equal to kept, rerun_from: report, to_infer: 0 and needs_inference: false.

The dry-run's needs_inference must agree with what sieval run then launches, for each outcome.