Fix skipped `test_model_xattn_padding` test by updating mask padding logic

[PrakshaaleJain]

Title:

Fix skipped test_model_xattn_padding test by updating mask padding logic

Description:

Currently, there is a skipped test in tests/models/unets/test_models_unet_2d_condition.py regarding cross-attention mask padding (test_model_xattn_padding).

The test is currently skipped with the following reason:

"we currently pad mask by target_length tokens (what unclip needs), whereas stable-diffusion's cross-attn needs to instead pad by remaining_length."

This is reflected by a dangling TODO in src/diffusers/models/attention_processor.py inside prepare_attention_mask (around line 740):

# TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
#       we want to instead pad by (0, remaining_length), where remaining_length is:
#       remaining_length: int = target_length - current_length
# TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)

Expected Behavior / Proposed Solution:

1. Update prepare_attention_mask in attention_processor.py (and any other relevant attention.py files) to pad the mask by remaining_length = target_length - current_length.
2. Ensure the MPS torch.zeros() hack is also updated to pad by remaining_length rather than target_length.
3. Verify that these changes do not break existing unclip tests/pipelines.
4. Remove @pytest.mark.skip from test_model_xattn_padding in tests/models/unets/test_models_unet_2d_condition.py to re-enable it.

Relevant Files:

• src/diffusers/models/attention_processor.py
• tests/models/unets/test_models_unet_2d_condition.py

@sayakpaul If this is indeed technical debt that needs fixing, I would be happy to put together a PR to update the logic and re-enable the test myself!