Operation Green CI II (#43537)
* Widen tolerances for generation a little
* Expand tols for MoEs only
* make fix-repo
* Factor it out into a function with a clearer name
* Factor it out into a function with a clearer name
* Factor it out into a function with a clearer name
* Stop swallowing errors in setUpClass
* Change cpm-ant default model (fixes capitalization in config.json)
* Do the same for integration tests
* Fix a source of test order flakiness for seamless_m4t
* Deepcopy processor kwargs so we don't mutate nested dicts
* Use a proper context manager
* make fix-repo
* More test state leaking
* Just about everything for PLBart is flaky
* Fix values for test_processor_glm46v.py. I wonder why this wasn't failing in the CI?
* Same fix for glm4v
* tokenization_auto fixes
* make fix-repo
* Update repo pointer for cpm-ant-testing
* Mark some of the big CB tests as slow for now
* Mark the CB tests as require_torch_accelerator
* Incorporate fixes from #43605
* Revert changes now we've incorporated Arthur/Ita's PR
