TLS certificates page: count reads upfront and keep partial ACME rows

[stephdl]

Targets sdl-8099, the branch of #1257, not main.

Description

A review pass on #1258 (HTTP routes page) raised two issues that apply here as well, since both pages share the same read pattern. This PR brings the two branches to the same shape, so the shared mixin planned as a follow-up on #1257 has one behaviour to extract instead of two.

Counters are set before the loop. listCertificates() and getAcmeServer() incremented inside the loop and decremented through a helper clamped with Math.max(0, n - 1). Two problems hide behind that clamp. The counter reaches zero between two iterations, so loadingCertificates / loadingServers can go false while the next createModuleTaskForApp is still awaited — the table drops out of its loading state with an empty or partial list. And a surplus decrement is swallowed rather than made impossible, which is what let the counter go out of balance in the first place. The counter is now this.traefikInstances.length before the loop and the decrements are plain --. Nothing is left to clamp: clearTaskListeners() already $offs the handlers of the previous round, so an event delivered late reaches nothing.

listCertificatesNum is released when the instance lookup is aborted. listInstalledModulesAborted reset loading.listInstalledModules but left the certificates counter untouched. On a reconnection the watcher restarts from listInstalledModules(), so if that task aborts the certificates chain never runs and whatever the counter held stays there — skeleton rows for good, with no way out but a manual reload.

One offline node no longer hides the others on the ACME tab. error.getAcmeServer is a single shared string, and NsDataTable renders its error notification instead of the rows (v-if="isErrorShown" … v-else-if="isLoading" … in NsDataTable.vue). On a cluster where one node is unreachable, the settings of every node that answered were replaced by the banner. The notification now renders above the table when there are rows to keep, and isErrorShown is left for the case where there is genuinely nothing to display. The Certificates tab already handles this through offlineTraefikInstances and listCertificatesErrors, so it needs no change.

Commits

1. fix(ui): count ACME and certificate reads upfront
2. fix(ui): keep the ACME rows readable when one node fails

Testing

core/ui has no test runner, so this was checked with lint and a production build, on each commit. Both pass.

To verify on a cluster:

• Both tables populate normally on a cold load and after a tab switch
• Save ACME settings on the connected node: the websocket drops, reconnects, and both tables refresh on their own with no skeleton rows left
• Reload during the Traefik restart and let it settle: neither table stays on skeleton rows
• Stop one traefik, open the ACME tab: the rows of the healthy nodes are readable, with the error notification above the table
• Restart it and save on another node: the missing row comes back and the notification clears
• Stop the leader so list-installed-modules aborts after a reconnection: the certificates table shows its error instead of loading for ever
• No duplicate rows in either table after several reconnections