Fix t5 failures (#43374)
* set eos_token_id in GenerationConfig
* fix expectation
* pop wo from dtype_plan when for bfloat16
* fix another expectation (same for amd & nvidia)
* last one
* fix repo-consistency
* fix
* override post_init() instead of __init__()
* use _keep_in_fp32_modules in from_pretrained() when dtype=torch.float16
* _build_dtype_plan_for_loading
* some fixes
* reverting
