Add missing dataset download to Qwen-Image validate examples

[rakhimovv]

The validate_full and validate_lora scripts under examples/qwen_image/model_training/ read data/example_image_dataset/ but never fetch it.

For edit/ and eligen/ this is not recoverable by running things in a particular order: at 4dbf980, nothing under examples/qwen_image/ downloads either subdirectory. So these fail even under the ordering in examples/dev_tools/unit_test.py, which runs model_inference before validate_lora. Running test_qwen_image() on a clean checkout, 7 of the 18 scripts in validate_lora fail like this:

Traceback (most recent call last):
  File ".../validate_lora/FireRed-Image-Edit-1.0.py", line 20, in <module>
    Image.open("data/example_image_dataset/edit/image1.jpg").resize((1024, 1024)),
FileNotFoundError: [Errno 2] No such file or directory: 'data/example_image_dataset/edit/image1.jpg'

(edit/image1.jpg is fetched by examples/boogu_image/, so a full cross-family run rescues some of these; edit/image_color.jpg and eligen/ are fetched nowhere in the repo.)

How the 18 validate_lora scripts divide:

• 11 read the dataset without fetching it. Of those, 7 fail as above, and 4 (canny, two layer, layer_v2) happen to pass because the matching model_inference scripts fetch that data into the same directory first.
• 3 already call dataset_snapshot_download (the Blockwise ControlNet scripts).
• 4 don't reference the dataset at all.

validate_full/ has the same problem in 7 more files. The sibling model_training/lora/*.sh scripts don't help either way: they fetch diffsynth_example_dataset to a different path.

This adds the call the three existing scripts use, following their convention of naming a single file directly (canny/image_1.jpg, depth/image_1.jpg) and globbing only where several are read (inpaint/*.jpg):

dataset_snapshot_download(
    dataset_id="DiffSynth-Studio/example_image_dataset",
    local_dir="./data/example_image_dataset",
    allow_file_pattern="edit/*.jpg"
)

Qwen-Image-Layered-Control-V2.py is included because it imports dataset_snapshot_download but never calls it, so a grep for the name makes it look fine.

This may also be what #1009 is asking about.

Checking

All 11 validate_lora scripts were run on 1×H100 against LoRAs trained by the model_training/lora/*.sh scripts in the same tree, and all produced output (e.g. Qwen-Image-Edit.py 80s, Qwen-Image-Edit-2509.py 105s, FireRed-Image-Edit-1.0.py 124s, Qwen-Image-EliGen.py 50s).

To confirm the download patterns themselves rather than relying on data an earlier script had left behind, I re-ran four of them in a fresh working directory with an empty data/: Qwen-Image-Edit.py (edit/image1.jpg), FireRed-Image-Edit-1.0.py (edit/*.jpg), Qwen-Image-EliGen.py (eligen/*.png) and Qwen-Image-Layered-Control.py (layer/image.png). Each fetched exactly its own inputs and produced an image, covering both pattern forms.

I have not run the 7 validate_full scripts — that needs a full fine-tune, and the accelerate configs are set up for 8 GPUs. Their change is the same one-call addition, but please treat those as unverified.

The same pattern appears in flux, wanvideo, z_image, mova and ltx2 (105 files repo-wide by the same static check, of which I have run 12). I kept this PR to qwen_image rather than sending a cross-family change unasked; glad to follow up on the others if that would help.