TST Run fast PEFT tests in normal CI (#45679)
PEFT tests were marked as slow, even though they're not slow. According
to a comment there, this was probably only done to avoid running the
tests before PEFT was released.
The @slow marker is now removed. All these tests use tiny models and on
my machine, with warm HF cache, they passed in <12 sec on CPU without
parallelism. It should thus be safe to assume that they are indeed fast
enough. With these tests now running in normal CI, we should be able to
prevent PEFT regressions in the future.
Note that the hotswapping tests are still marked as slow, as they partly
require calling torch.compile.
