fix: fallback adapters report the active instance's model/provider

[Fyko]

What

STT/TTS/LLM/realtime fallback adapters now report the model/provider of the underlying instance that most recently served a request, instead of the hardcoded "FallbackAdapter"/"livekit".

Why

Turn metrics and spans read model/provider off the top-level component, so with a fallback adapter every lk.agents.turn.* histogram collapses into one model_name="FallbackAdapter" series — you can't tell your primary's ttfb from your fallback's. Discussed with the LiveKit team in slack; surfacing the actual provider was agreed on.

Semantics: the primary is reported before any traffic, the last instance to produce output after that. Background recovery probes never relabel — the caller never saw their output. The serving streams mark the active instance directly (same delegation pattern as stt.StreamAdapter and the realtime adapter's capabilities), no event listening needed.

Tests added for all four adapters, including probe-doesn't-relabel and label-follows-recovery cases.

[CLAassistant]

All committers have signed the CLA.

[longcw]

The nested llm_request span already carries the real gen_ai.request.model, so model can stay "FallbackAdapter".

The only gap is lk.agents.turn.*, whose labels come from MetricsReport at three sites (agent_activity:3380, :3385, :4457). Maybe add a single metrics_metadata property, read at those three sites, splits the histograms and keeps model, ErrorEvent, and the user_turn span stable.

ifTYPE_CHECKING:
from ..llm.chat_contextimportMetricsMetadata@propertydefmetrics_metadata(self) ->MetricsMetadata: # base classes return their own model/providerreturnself._active_instance.metrics_metadata