Attach real input/output/token usage to provider spans for Langfuse cost tracking

Langfuse showed $0.00 cost on every trace even after tracing was wired up:
provider.bedrock_converse.* spans carried only latency, no content or usage.
Converse's own response already includes token usage -- it just wasn't being
read. Captured on the adapter (last_usage, no interface change) and attached
to the span using Langfuse's own OTel attribute convention
(langfuse.observation.*), which its docs say takes precedence over generic
GenAI attributes. Verified live: a fresh trace now shows real prompt/completion
tokens and Langfuse computes real cost once given the model's actual per-token
price.

apps/ai-service/adapters/bedrock_converse.py

@@ -78,6 +78,10 @@ class BedrockConverseAnswerGenerator:
         self._client = client
         self._model_id = model_id
         self._max_tokens = max_tokens
+        # Converse's own response usage, captured for the observability layer
+        # (rag/instrumentation.py) to attach to its span. Not part of the
+        # `generate()` return contract -- every existing caller is unaffected.
+        self.last_usage: dict[str, int] | None = None
 
     @property
     def model_id(self) -> str:
@@ -126,6 +130,8 @@ class BedrockConverseAnswerGenerator:
                 f"{self._model_id} could not be invoked: {type(error).__name__}"
             ) from error
 
+        self.last_usage = dict(response.get("usage") or {}) or None
+
         if response.get("stopReason") in _EMPTY_STOP_REASONS:
             raise AnswerGenerationUnavailable(
                 f"{self._model_id} produced no usable content "