[train_instruct_pix2pix_sdxl.py] Fix the LR scheduler when num_train_…

[AayushMainali-Github]

What does this PR do?

Fixes the learning-rate scheduler setup in examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py when training is launched with --num_train_epochs in a distributed environment.

This is one script from the checklist in #8384. The scheduler was previously created from the unsharded dataloader length, so num_training_steps did not match the real optimization step count after accelerator.prepare. The change matches the #8312 pattern already applied to train_instruct_pix2pix.py.

Fixes #8384

Minimal training command using --num_train_epochs:

accelerate launch examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py \
  --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-xl-pipe \
  --dataset_name hf-internal-testing/instructpix2pix-10-samples \
  --resolution 64 \
  --random_flip \
  --train_batch_size 1 \
  --num_train_epochs 1 \
  --lr_warmup_steps 0 \
  --output_dir /tmp/instruct-pix2pix-sdxl-epochs \
  --seed 0