Add opt-in Exclusive Self Attention support

[taivu1998]

Summary

Adds opt-in support for Exclusive Self Attention (XSA) in the shared Diffusers attention stack.

• Adds exclusive_self_attention=False to Attention.
• Applies the XSA projection after attention output and before output projection for verified self-attention paths.
• Wires the option through BasicTransformerBlock.
• Exposes the option through Transformer2DModel, DiTTransformer2DModel, and PixArtTransformer2DModel configs.
• Adds focused tests for the XSA math, self-vs-cross attention gating, processor variants, and model config/forward propagation.

Fixes #13447.

Motivation

Issue #13447 requests an optional Exclusive Self Attention mode based on:

z_i = y_i - (y_i @ v_i) / (v_i @ v_i) * v_i

This PR implements the equivalent normalized projection form:

value_normalized = F.normalize(value, p=2, dim=-1)
hidden_states = hidden_states - (hidden_states * value_normalized).sum(dim=-1, keepdim=True) * value_normalized

The flag is stored on the Attention module rather than on processor instances, so it remains stable when attention processors are swapped.

Scope

Supported in this PR:

• AttnProcessor
• AttnProcessor2_0
• FusedAttnProcessor2_0
• SlicedAttnProcessor
• XFormersAttnProcessor
• AttnProcessorNPU
• XLAFlashAttnProcessor2_0

The projection is applied only when the current call is true self-attention (encoder_hidden_states is None). Cross-attention remains unchanged even if sequence lengths happen to match.

Model-level exposure is included for:

• Transformer2DModel
• DiTTransformer2DModel
• PixArtTransformer2DModel

Deferred

This intentionally does not change added-KV processors, joint attention processors, direct-dispatch model-specific processors, or full U-Net constructor propagation. Those paths have more ambiguous token/value pairing semantics and are better handled in follow-up PRs if maintainers want broader coverage.

Validation

Ran:

PYTHONPATH=src python -m pytest tests/models/test_exclusive_self_attention.py
PYTHONPATH=src python -m pytest tests/models/test_layers_utils.py -k "test_spatial_transformer_default or exclusive_self_attention"
PYTHONPATH=src python -m pytest tests/models/transformers/test_models_dit_transformer2d.py -k "test_output or exclusive_self_attention"
PYTHONPATH=src python -m pytest tests/models/transformers/test_models_pixart_transformer2d.py -k "test_output or exclusive_self_attention"
python -m py_compile src/diffusers/models/attention_processor.py src/diffusers/models/attention.py src/diffusers/models/transformers/transformer_2d.py src/diffusers/models/transformers/dit_transformer_2d.py src/diffusers/models/transformers/pixart_transformer_2d.py tests/models/test_exclusive_self_attention.py tests/models/test_layers_utils.py tests/models/transformers/test_models_dit_transformer2d.py tests/models/transformers/test_models_pixart_transformer2d.py
uvx ruff check src/diffusers/models/attention_processor.py src/diffusers/models/attention.py src/diffusers/models/transformers/transformer_2d.py src/diffusers/models/transformers/dit_transformer_2d.py src/diffusers/models/transformers/pixart_transformer_2d.py tests/models/test_exclusive_self_attention.py tests/models/test_layers_utils.py tests/models/transformers/test_models_dit_transformer2d.py tests/models/transformers/test_models_pixart_transformer2d.py
uvx ruff format --check src/diffusers/models/attention_processor.py src/diffusers/models/attention.py src/diffusers/models/transformers/transformer_2d.py src/diffusers/models/transformers/dit_transformer_2d.py src/diffusers/models/transformers/pixart_transformer_2d.py tests/models/test_exclusive_self_attention.py tests/models/test_layers_utils.py tests/models/transformers/test_models_dit_transformer2d.py tests/models/transformers/test_models_pixart_transformer2d.py
git diff --check