fix(grafana): stop short writer gaps from paging false alerts (#1695)

[groupsky]

Fixes #1695.

Why

A short gap in InfluxDB data (for example a deploy that restarts mqtt-influx) paged false alerts:

• 8 rules used noDataState: NoData. Grafana 9.5 sends NoData at once as a "DATA SOURCE ISSUE" page and does not wait for for (routy: deploy restarting influxdb + grafana sends false "DATA SOURCE ISSUE" pages — grafana evaluates before influxdb is up #1682).
• webhook-totalout01 (Total Power Outage) uses no data as its blackout signal, with for: 1m. current_power has one writer, mqtt-influx, so a writer gap of about 2.5 min looked like a blackout.

What changed

noDataState for the 8 rules (each has a short comment with the reason):

uid	new	why
boiler-temperature-sensor-failure	Alerting, for 30m → 5m	absence rule: the count gives no rows when data stops, so no data is the signal
sunseeker-connection-lost	Alerting, for 30m → 5m	same
boiler-excessive-power-consumption	OK	high-power rule: no data = no high reading
webhook-lA5nSTt4k (heat pump overload)	OK	same
webhook-lA5nSTtVk (main power)	OK	same
webhook-z157Sot4z (water pump)	OK	same
webhook-rAc7STtVk (dishwasher done)	OK	reminder: no data = no finished cycle
webhook-XA5nIopVk (laundry done)	OK	same

For the 2 absence rules, the 30 min query window already does the waiting. With Alerting, for starts only after the window is empty, so for: 30m would page 30 min later than before. With for: 5m they page about 35 min after data stops (before: about 30 min, as a "DATA SOURCE ISSUE"). A short writer gap cannot empty a 30 min window. Both have execErrState: Alerting, so a query error now pages after 5 min instead of 30; an InfluxDB restart takes about 90 s.

Before this change all 8 were Normal on the production host, so Alerting does not page on deploy.

Total Power Outage: for: 1m → for: 5m.

Trade-off: a writer gap of up to 5 min is silent. A real blackout now pages about 4 min later: about 6.5–7.5 min after the meters stop (90 s query window + 5 min for + up to 1 min to the next evaluation), instead of about 2.5–3.5 min. A real outage still pages. I chose this over a "writer is alive" check because it is one line and needs no new data source.

CI check. validate-grafana-alerts now fails when a provisioned rule uses noDataState: NoData, or leaves it unset (Grafana 9.5 rules_types.go defaults a blank value to NoData). There are tests for both, for .json files, and for the real directory.

Test

• npm test: 155 passed. Before the rule edits, the real-directory test failed and listed the 8 rules.
• npm run validate: all checks pass.
• After deploy: restart mqtt-influx and check the Telegram send log. The result goes in the issue.