📊 telemetry: publish round progress from the foreground path too

[george-connito]

validator_round_miners_{scored,failed,pending} were written only by BackgroundEvalWorker._record_metrics, whose loop starts when the eval window opens at Merge. Foreground evals run earlier, during Submission, and accumulate in scored_uids with nothing publishing them — so the dashboard's "Evaluated N of M" panel showed the previous round's final count for the first ~14 minutes of every round, then jumped straight to the foreground total (measured on cycle 16674: Submission 21:32, panel flipped at Merge 21:46 with a first value of 4).

The counts were always correct — evaluate_foreground_round does call mark_scored, and Round.stats() reads whole-round state. Only the publish was background-scoped.

Lift the publish out of BackgroundEvalWorker into telemetry.set_round_progress(round_id, scored, failed, pending) + Round.publish_progress(), and call it from:

• Round.freeze (via run.py, next to the lifecycle-step-0 publish) so the round appears in the metric at freeze with scored=0;
• each foreground mark_scored / mark_failed / mark_validation_failed, matching the background worker's existing mark → publish → prune ordering;
• the background worker, now via the shared method.

note_round_series moves into the shared helper, so the labelsets are still evicted on the normal cutoff. The vestigial scored_inc argument (never read) is dropped. The download worker's partial pending/failed publishes are left untouched — out of scope and still correct.

Tests: new test_round_progress_publish.py covers the shared helper (all three gauges, eviction registration, never-raises), freeze-time seeding, the foreground regression (mark_scored → counter moves with no background worker), failure/validation-failure counting, stats parity, lock-safety, and per-round isolation. Full round/telemetry suite (87) green in the stable image.