Pin protobuf<6 for ray[serve]==2.55.0

ray[serve]==2.55.0's _proto_to_dict reads FieldDescriptor.label, which
protobuf>=6 removed, breaking beginner_source/serving_tutorial.py with
"'google._upb._message.FieldDescriptor' object has no attribute 'label'".
The fresh CI image (rebuilt by the torch bump) pulled protobuf 6. ray 2.55.0
declares protobuf>=3.20.3 with no upper bound, so this pin is compatible.
Unrelated to torch 2.13.

Co-authored-by: Claude

Author: atalman

.ci/docker/requirements.txt

@@ -37,6 +37,8 @@ setuptools<81
 awscliv2==2.1.1
 flask
 spacy==3.7.1  # Keep this version consistent with the model version in .jenkins/build.sh
+# Pin protobuf: ray[serve]==2.55.0 uses FieldDescriptor.label, removed in protobuf>=6.
+protobuf<6
 ray[serve,train,tune]==2.55.0
 tiktoken
 tensorboard