[hf-diffusers] Use the latest transformers for diffusers (#1027)

* Skip saving pretrained model for diffuser.

* Skip saving pretrained model for diffuser, in nightly

* Skip checkpointing in hf-diffusers, hf-glue conv tests.

* Use the latest transformers

tests/pytorch/nightly/hf-diffusers.libsonnet

@@ -54,6 +54,7 @@ local tpus = import 'templates/tpus.libsonnet';
         cd examples/text_to_image
         sed '/accelerate/d' requirements.txt > clean_requirements.txt
         sed '/torchvision/d' requirements.txt > clean_requirements.txt
+        sed -i 's/transformers>=.*/transformers>=4.26.0/g' clean_requirements.txt
         pip install -r clean_requirements.txt
 
         # Skip saving the pretrained model, which contains invalid tensor storage