s57_layer: invalidate cached tide_offset when chart_datum disappears (datum_source=none)

[rolker]

Context

Follow-up surfaced during the pre-push review of rolker/mru_transform#25
(datum resolution chain off-VDatum).

s57_layer caches tide_offset_ and only updates it when a fresh
chart_datum → sea_surface lookup succeeds and differs by
tide_invalidate_threshold_. When the lookup throws (the chart_datum frame is
no longer broadcast / has expired from the TF buffer), it logs a throttled WARN
and keeps the last tide_offset_:

try {
  auto transform = tf_->lookupTransform(chart_datum_frame_, sea_surface_frame_, ...);
  ...
} catch (const tf2::TransformException& e) {
  // throttled WARN, tide_offset_ unchanged
}

Before #25, off-coverage simply meant chart_datum was never published, so
tide_offset_ would never have been set. After #25, chart_datum is published
in more places (VDatum, polygon config, lake_datum param) and can legitimately
disappear while underway when the boat moves into an area where nothing
matches (the node intentionally stops broadcasting chart_datum and publishes
datum_source == "none"). So the costmap can silently keep correcting against a
stale offset from a different region — a newly-reachable "stale data" path.

Proposal

Treat datum disappearance as a first-class transition:

• Subscribe to the datum_source topic (std_msgs/String, latched) published by
  chart_datum_node. When it reports "none", clear/invalidate tide_offset_
  (and ideally invalidate the tide-corrected layer) rather than holding the last
  value.
• Alternatively, treat a sustained lookup failure (frame absent beyond a timeout)
  as invalidation.

Acceptance

• tide_offset_ is cleared/invalidated when the datum becomes unavailable
  (datum_source == "none" or sustained lookup failure)
• No regression when the datum is continuously available

Related: rolker/mru_transform#25, rolker/mru_transform#27.

---

Authored-By: Claude Code Agent
Model: Claude Opus 4.8 (1M context)