allow strict=True in load_checkpoint_and_dispatch

[yiyixuxu]

System Info

accelerate = 0.29.1

Information

this PR #2588 breaks the loading behavior in diffusers when:

1. accelerate.load_checkpoint_and_dispatch is used
2. the checkpoint contains deprecated weight names for the attention class

As you can see here, we rely on the error to rename the parameters' names on our end. Is there a better way to do this in diffusers?
The alternative is to allow us a pass a strict=True from load_checkpoint_and_dispatch- let me know what you think :)

Reproduction

this script fails currently but it should not

from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
    "hf-internal-testing/tiny-stable-diffusion-pipe", device_map="sequential", safety_checker=None
)

[yiyixuxu]

cc @SunMarc

[SunMarc]

Let's just pass strict=True in load_checkpoint_and_dispatch. Sorry for introducing a breaking change ;)