feat(litellm): trace transcription calls

Instrument litellm.transcription() and litellm.atranscription() with
Braintrust spans, including audio input attachments and transcription
text output.

Add focused VCR-backed regression coverage for the sync and async
transcription paths.

Author: AbhiPrasad

py/src/braintrust/integrations/litellm/__init__.py

@@ -9,8 +9,9 @@ def patch_litellm() -> bool:
 
     This wraps litellm.completion, litellm.acompletion, litellm.responses,
     litellm.aresponses, litellm.image_generation, litellm.aimage_generation,
-    litellm.embedding, litellm.aembedding, and litellm.moderation to
-    automatically create Braintrust spans with detailed token metrics,
+    litellm.embedding, litellm.aembedding, litellm.moderation,
+    litellm.transcription, and litellm.atranscription to automatically
+    create Braintrust spans with detailed token metrics,
     timing, and costs.
 
     Returns: