[tests] refactor ltx2 autoencoder tests to use latest mixins

[sayakpaul]

As the PR title says.

[dg845]

Thanks for the PR! If I run the tests locally, the following tests fail for both the LTX-2.X video and audio VAEs (and on the CI):

1. test_from_save_pretrained
2. test_from_save_pretrained_variant
3. test_determinism
4. test_sharded_checkpoints_with_parallel_loading

Since test_determinism is one of the tests that fail, I wonder if the underlying non-determinism is also affecting the save-then-load tests as well.

I also see that the following tests fail locally (but not in the above CI run):

1. test_sharded_checkpoints
2. test_sharded_checkpoints_with_variant

Finally, the test_group_offloading_with_disk[leaf_level-False] test also fails locally for the LTX-2.X audio VAE, so it seems like it might have some offloading issues. Could you look into these test failures?

[sayakpaul]

@dg845 thanks for testing, requested for a review a bit sooner. I think it's good now.

[dg845]

Thanks! I see that the TestAutoencoderKLLTX2AudioMemory::test_group_offloading_with_disk[leaf_level-False] test still fails locally for the LTX-2 audio VAE, do you think we should skip this particular test?

[dg845]

Actually, it seems like the LTX-2 audio VAE test_group_offloading_with_disk[leaf_level-False] is flaky, since sometimes it passes and sometimes it fails. Not sure why.

[sayakpaul]

We can add is_flaky to the test to reflect that it's flaky?