Fast LoRA initialization by skipping redundant linearizations

[isidentical]

What does this PR do?

Part of #4975, along with #4976 and #4979. This PR leverages the accelerate's init empty weights (just like diffusers automatically does on certain models, like vae for SD) when we are initializing the LoRA weights for the first them. This grants a great boost for avoiding unnecessary computations which would be overriden just a moment later. On top of #4976 and #4979, this PR makes the load_lora_weights() + unload_lora_weights() cycle take around a second (1.05seconds to be precise) from ~6 seconds without any of them. Relative effect on top of the first two PRs (combined) is about 1.33x.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[patrickvonplaten]

I think this is already fixed by: #4994

[isidentical]

Yep, #4994 fixes this!