Upgrade neuron to 2.20.2

This change also fixed loading Neuron models. Neuron uses old pytorch,
 before the change I was getting

TypeError: infer_schema() takes 1 positional argument but 2 were given

which was caused by custom_op direct registration. I'm disabling that
if custom_op is not available.

Dockerfile.neuron

@@ -1,5 +1,6 @@
 # default base image
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.0-ubuntu20.04"
+# https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.1.2-neuronx-py310-sdk2.20.2-ubuntu20.04"
 
 FROM $BASE_IMAGE
 

vllm/utils.py

@@ -1628,7 +1628,7 @@ def direct_register_custom_op(
     library object. If you want to bind the operator to a different library,
     make sure the library object is alive when the operator is used.
     """
-    if is_in_doc_build():
+    if is_in_doc_build() or not supports_custom_op():
         return
     import torch.library
     if hasattr(torch.library, "infer_schema"):