fix: tolerate transient poll failures before marking vehicle offline

[nanomad]

The vehicle available topic went offline on the first failed poll, so every HA entity flipped to unavailable for each "try again later" reply from the SAIC server. The next poll usually succeeds.

The vehicle is now marked offline only after 3 consecutive failed polls. A successful poll resets the count. Vehicles removed from the API list still go offline immediately. Polling backoff and the last-error timestamp are unchanged.

Why 3: I measured failure runs in about 220k logged status API attempts (Mar 2024 – Oct 2026, one vehicle, grouped into 91,918 polls, 3.2% failed). 82% of failure runs are a single failed poll.

Threshold	Offline flaps	Median delay before offline
1 (before)	995	0
2	175	1 min
3	109	2.5 min
5	85	9.6 min

3 removes 89% of the flaps. Going higher adds little and delays real outages. Caveats: one vehicle, and polls were grouped by a 20 s gap rather than from the gateway's own failure marks.

Closes #313