feat: add log-based CoderdPanic alert via the Loki ruler

[ibetitsmike]

Moves the CoderdPanic alert from dogfood-specific Loki values into the chart, as requested by @dannykopping in coder/dogfood#454 (review comment: https://github.com/coder/dogfood/pull/454#discussion_r3896503702).

What

Adds the chart's first log-based (Loki ruler) alerting rule:

• CoderdPanic fires when a coderd pod logs a fatal panic: / fatal error:, or recovers from an HTTP handler panic (panic serving http request), within the configured period (default 5m).
• New coder-logs-alerts ConfigMap (templates/configmap-loki-alerts.yaml + templates/_loki-alerts.tpl), labeled loki_rule: "" so the Loki rules sidecar loads it into the ruler. The ruler already ships alerts to Alertmanager (rulerConfig.alertmanager_url).
• Configured via global.coder.alerts.coderd.groups.Panics (enabled/period/thresholds), following the existing alert group conventions, with a runbook_url label and a new runbook section.

Notes on the matcher

• Panic output is untimestamped, so the collector's multiline stage folds tracebacks into the preceding timestamped log entry; (?m)^ matches the embedded panic line start.
• coderd.aibridge* loggers are excluded: they echo LLM chat content, which can quote Go tracebacks pasted by users (observed false positives on dev.coder.com).

Validation

• make lint passes (helm lint --strict, pint, helm-docs README regen, compiled/ regen with helm v3.17.1 + yq v4.42.1).
• The matcher was backtested against 7 days of live dev.coder.com Loki data in coder/dogfood#454 (100% true positives). The chart's coderd-selector-based variant was verified to return identical per-pod results to the backtested job-based selector on the same data.

Once released, dogfood will bump the chart version and keep only its Slack routing for the alert (coder/dogfood#454 is being reworked accordingly).

Disclosure: this PR was created by Xum (an AI agent) on behalf of @ibetitsmike.