fix: raise informative ValueError on architecture mismatch when loading LoRA weights (SD 1.5 into SDXL)

[Liauuu]

Description

This PR addresses and fixes Issue #11286, where loading an incompatible, older architecture LoRA (e.g., Stable Diffusion 1.5) into a newer pipeline (e.g., Stable Diffusion XL) triggers an unhandled RuntimeError: size mismatch deep within the tensor operation layers.

Instead of allowing the program to mutably inject adapter shells and crash abruptly midway through tensor addition, this fix introduces a proactive, defensive dimension validation step.

Key Changes

1. Added _validate_lora_weight_compatibility helper in src/diffusers/utils/peft_utils.py:
   • Inspects the incoming state_dict keys (lora_A and lora_B weights).
   • Dynamically fetches target submodules using model.get_submodule().
   • Cross-references incoming tensor shapes and ranks against expected base layer dimensions (in_features/in_channels and ndim).
2. Injected Validation Gateways:
   • Integrated inside PeftAdapterMixin.load_lora_adapter (src/diffusers/loaders/peft.py) before PEFT commits mutations via inject_adapter_in_model.
   • Integrated inside lora_base.py for text encoder pathways before executing text_encoder.load_adapter.
3. Informative Error Reporting:
   • Throws a clean, informative ValueError mapping out the precise architecture discrepancy (e.g., 4D Conv LoRA vs 2D Linear layers, or 768-dim context vs 2048-dim target dimensions) so users instantly recognize version mismatches.

Testing

• Added a full automated regression test suite inside tests/lora/test_lora_layers_sdxl.py: test_load_sd15_lora_into_sdxl_raises_incompatibility_error.
• The test generates a synthetic, cross-architecture incompatible state_dict and asserts the exact ValueError criteria and error strings.
• Local verification status: PASSED [100%] (Ran locally using pytest).

Fixes #11286