Cover the full metric set in the dashboard and add a shareable export

[Pixnop]

Closes #45.

Th3Dilli was right that the dashboard read as unfinished. It had eight panels
against a metric set several times that size, so most of what the mod measures
had nowhere to show up.

The dashboard now runs top to bottom as a story about the server. A glance strip
of the numbers you check first, then tick health with p50/p95/p99 out of the
histogram and busy time drawn against the budget, players and ping and deaths,
chunks and entities with the per-code breakdown stacked, worldgen, network on
both the engine's per-second families and the byte totals so TCP and UDP are
both visible, then suspends, engine warnings and log levels, and a runtime row
for the dotnet_* families. 33 data panels across 8 rows, up from 8 panels.

Panels whose families are conditional say so in their own description. Someone
staring at an empty graph should learn from the panel that it needs the engine
probe, or RuntimeMetrics, instead of concluding the mod is broken. The same
goes for the numbers that need interpreting: busy time meeting the budget is a
saturated server, and the suspend seconds are the freeze players actually feel
during an autosave.

pulse-overview-shared.json is the second ask. Same dashboard in
export-for-sharing form, so Grafana's import dialog asks which Prometheus to
bind rather than handing you 33 panels wired to the uid pulse-prom, which
exists only on a Grafana provisioned from this directory. It is generated by
make-shared.py rather than kept in sync by hand; the README says so, in the
hope that nobody edits it directly.

check-dashboard.py covers the failure mode with no error message. Overlapping
panels, a panel running past the 24 column grid and duplicate panel ids all get
drawn wrong or dropped without Grafana saying a word, which is a miserable thing
to chase by eye. Stdlib only, like the generator.

Checked rather than assumed:

• Both files pass check-dashboard.py, and the checker was itself run against a
  deliberately broken dashboard to confirm it catches all five violations rather
  than always printing ok.
• All 45 PromQL expressions parse, via promtool check rules over a generated
  rule file.
• Metric and label names come from the source, not from memory. The dotnet_*
  names in particular were taken by enumerating the real System.Runtime meter
  and pushing the names through the same mapping PrometheusText.MetricName
  applies, which is where dotnet_thread_pool_queue_length_total gets its
  _total.
• Grafana 13.2.0 provisioned the dashboard from these files and loaded all 41
  panels with no errors, and the shared file imported cleanly through
  /api/dashboards/import with every datasource reference resolved.
• Regenerating the shared file leaves the tree clean, so the committed copy is
  exactly what the script produces.

Left out on purpose: no template variables, since there is one server per
scrape target here and a datasource picker would just be one more thing to
get wrong on import. No alert rules either, contrib/alerts already owns those.