fix(mqtt): don't guess a DID that onboarding already paired

[Lash-L]

Follow-up to #165 (issue #163).

Problem

#165 made the DID/DUID pairs learned during onboarding authoritative in the topic bridge's DUID->DID map. The older unique-model fallback still runs first, though, and it can map an unpaired DUID to a DID that runtime_credentials.json already pairs with a different DUID.

Example: the inventory's only a168 isn't onboarded yet, and the only a168 DID with a known model belongs to another vacuum (for example, the one it replaced). On main the map becomes:

1mRc3mFQhMWxkAQIAGI46w -> 1103841353419   (model guess)
6aEI34L0ZtsALO8XtccuPw -> 1103841353419   (learned pair)

Commands for the unpaired vacuum are then sent to the other vacuum's topic instead of being dropped with the "no deterministic rr/d mapping" warning.

Fix

• Collect the learned pairs before the inventory loop, and skip the model fallback for any DID they already claim. Learned pairs still override everything else, as in fix(mqtt): map same-model vacuums via learned DID/DUID pairs #165.
• Treat an exception from runtime_credentials.devices() as "no learned pairs", the same way key_models_by_did() is handled just above it. Without this, an exception would drop the bridge's broker connection and send it through its reconnect loop.
• Remove the --bridge-device-* hint from the two drop warnings. That option no longer exists; the warning now says to onboard the vacuum.

Testing

• New test test_topic_bridge_model_fallback_does_not_reuse_did_learned_for_another_duid. It fails on current main and passes with this change.
• Full suite: 335 passed.