Add DiffSynth blockwise ControlNet support to QwenImageControlNetModel

[aryabyte21]

Summary

Adds support for the DiffSynth-Studio blockwise ControlNet architecture to the existing QwenImageControlNetModel, addressing the maintainer's feedback on PR #12317 by integrating it into the existing class rather than creating a separate one.

What changed

• Added controlnet_block_type config parameter to QwenImageControlNetModel ("linear" default for InstantX, "blockwise" for DiffSynth)
• Added BlockWiseControlBlock module: RMSNorm + MLP fusion block that normalizes both base hidden states and control features separately before fusing via Linear → GELU → Linear (zero-initialized output)
• No pipeline changes needed — both variants produce the same controlnet_block_samples output format
• Added conversion script for DiffSynth checkpoints (scripts/convert_diffsynth_blockwise_controlnet_to_diffusers.py)

Key design decision

Per @yiyixuxu's review comment on #12317:

"is this the only difference between Diffsynth & instantX controlnet, can we just add the change to the existing one?"

Yes — the only architectural difference is in the controlnet blocks themselves. The blockwise variant uses BlockWiseControlBlock(x, y) (two inputs: base + control features) instead of zero_module(nn.Linear)(x) (one input). Everything else — the pipeline, transformer integration, and output format — is identical.

Available checkpoints

• DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Canny
• DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Depth
• DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Inpaint

Usage

from diffusers import QwenImageControlNetModel

# Blockwise variant (DiffSynth)
controlnet = QwenImageControlNetModel.from_pretrained(
    "path/to/converted/blockwise-controlnet"
)

# Or create from base transformer
controlnet = QwenImageControlNetModel.from_transformer(
    transformer,
    controlnet_block_type="blockwise",
)

Closes #12221

Test plan

• Existing test_qwen_controlnet passes (no regression)
• Existing test_qwen_controlnet_multicondition passes (no regression)
• New test_qwen_blockwise_controlnet passes
• New test_qwen_blockwise_controlnet_from_transformer passes
• Full GPU inference test with converted DiffSynth checkpoint (needs GPU)