fix(monitoring): harden ONT/CM/Cellular Disable/Enable toggle (review follow-ups)

[Optic00]

Pattern-wide robustness follow-ups to the device-monitoring Disable/Enable toggle (#1044 ONT, #1046 CM, #1047 Cellular), applied across all three device types. From an adversarial review of the toggle; two of these (#1, #4) were flagged as good catches.

#1 - Make the anti-resurrection guard atomic

UpdateXxxPollResultAsync read Enabled and then called SaveChanges separately - a TOCTOU where a Disable committing in between let an in-flight poll still overwrite the paused row's frozen LastError/LastPolled. It now does a single conditional ExecuteUpdate(...).Where(x => x.Id == id && x.Enabled) (returns rows > 0), so the check and the write are one atomic statement. Repo tests moved from the EF InMemory provider to SQLite in-memory (InMemory can't execute ExecuteUpdate), plus a relational mid-poll-disable regression per type.

#2 - Don't contact / falsely report a disabled device on the manual + dashboard paths

The public poll entry points fetched the device before the guard ran, so refreshing a disabled device still hit the network - and Cellular returned (true, "polled successfully") while nothing persisted. PollCmAsync, PollModemAsync, and the ONT manual poll now short-circuit when the config is disabled (Cellular returns a clear "modem is disabled" message), and the CM/Cellular/ONT stats panels skip auto-polling a disabled config. The non-persisting Test/Probe buttons are unchanged.

#3 - One control for enable/disable (UX change)

Each edit form still had its own "Enable polling" checkbox, competing with the new row toggle - open the edit form, toggle the row off, click Update, and the stale form value re-enabled it. Removed that checkbox from the ONT/CM/Cellular edit forms (the row Disable/Enable button is now the single control) and stopped SaveXxxConfigurationAsync from writing Enabled on update, so SetXxxEnabledAsync is the sole writer. Starlink's edit form is untouched (it has no toggle). New tests assert Save preserves the DB Enabled value.

#4 - A persistence exception is no longer mistaken for "disabled"

The CM/Cellular poll success helpers caught every repository exception and returned false - the same value the guard uses for a disabled row - so a transient DB write failure on an enabled device silently dropped valid stats (and Cellular reported success). They now rethrow, so a real failure surfaces via the poll error path; false only ever means "disabled" (matching how ONT already behaved).

Tests

NetworkOptimizer.Storage.Tests 183/183 and NetworkOptimizer.Web.Tests 933/933 green; solution builds clean (0 warnings).

Off dev after #1044/#1046/#1047 merged.

[Optic00]

Moving this to draft / on hold — no rush, syncing with what's already in dev.

• Update dependency Lextm.SharpSnmpLib to 12.5.7 - autoclosed #2 here (don't contact / false-report disabled devices on the manual + dashboard paths) is already covered by 618c2e90 ("Dashboard ONT/CM/Cellular cards: respect Device Monitoring disable toggle"), so that part will be dropped — it was duplicate work.
• dev has since moved (Starlink toggle, etc.), so this needs a rebase and currently conflicts.

Plan for the next dev cycle: rebase onto fresh dev, drop the duplicated #2, resolve the remaining #1 (atomic guard) / #3 (single Enable control) / #4 (exception handling) conflicts, and likely extend #1/#3/#4 to Starlink too now that it has the same toggle — so the hardening is pattern-wide across all four types.