chore(replay): improve replay robustness with un-validated construction (#3414)

# What does this PR do?

some providers do not produce spec compliant outputs. when this happens
the replay infra will fail to construct the proper types and will return
a dict to the client. the client likely does not expect a dict.

this was discovered with tgi, which returns finish_reason="" when valid
values are "stop", "length" or "content_filter"

## Test Plan

ci

Author: mattf

llama_stack/testing/inference_recorder.py

@@ -105,8 +105,12 @@ def _deserialize_response(data: dict[str, Any]) -> Any:
 
             return cls.model_validate(data["__data__"])
         except (ImportError, AttributeError, TypeError, ValueError) as e:
-            logger.warning(f"Failed to deserialize object of type {data['__type__']}: {e}")
-            return data["__data__"]
+            logger.warning(f"Failed to deserialize object of type {data['__type__']} with model_validate: {e}")
+            try:
+                return cls.model_construct(**data["__data__"])
+            except Exception as e:
+                logger.warning(f"Failed to deserialize object of type {data['__type__']} with model_construct: {e}")
+                return data["__data__"]
 
     return data