fix(inference): propagate 401 errors

Remote provider authentication errors (401/403) were being converted to 500 Internal Server Error, hiding the real cause from users.

Now checks if exceptions have a status_code attribute and preserves it. This fixes authentication error handling for all remote inference providers using OpenAI SDK (groq, openai, together, fireworks, etc.) and similar provider SDKs.

Before:
- HTTP 500: "Internal server error: An unexpected error occurred."

After:
- HTTP 401: "Error code: 401 - Invalid API Key"

Fixes #2990

Test Plan:
1. Build stack: llama stack build --image-type venv --providers inference=remote::groq
2. Start stack: llama stack run
3. Send request with invalid API key via x-llamastack-provider-data header
4. Verify response is 401 with provider error message (not 500)
5. Repeat for openai, together, fireworks providers

llama_stack/core/server/server.py

@@ -138,6 +138,13 @@ def translate_exception(exc: Exception) -> HTTPException | RequestValidationErro
         return HTTPException(status_code=httpx.codes.NOT_IMPLEMENTED, detail=f"Not implemented: {str(exc)}")
     elif isinstance(exc, AuthenticationRequiredError):
         return HTTPException(status_code=httpx.codes.UNAUTHORIZED, detail=f"Authentication required: {str(exc)}")
+    elif hasattr(exc, "status_code") and isinstance(getattr(exc, "status_code", None), int):
+        # Handle provider SDK exceptions (e.g., OpenAI's APIStatusError and subclasses)
+        # These include AuthenticationError (401), PermissionDeniedError (403), etc.
+        # This preserves the actual HTTP status code from the provider
+        status_code = exc.status_code
+        detail = str(exc)
+        return HTTPException(status_code=status_code, detail=detail)
     else:
         return HTTPException(
             status_code=httpx.codes.INTERNAL_SERVER_ERROR,