[train_t2i_adapter_sdxl.py] Fix the LR scheduler when num_train_epochs is passed in a distributed training env

[adi-IL]

Fixes #8384

Follow-up to #14527 and #14528, still one script. This PR updates examples/t2i_adapter/train_t2i_adapter_sdxl.py.

What was wrong

When --num_train_epochs is used, train_t2i_adapter_sdxl.py built the LR schedule from the unsharded dataloader and ignored accelerator.num_processes, resulting in mismatched warmup and total training steps across distributed environments.

What changed

Applies the #8312 pattern already established across the DreamBooth, ControlNet, and Text-to-Image trainers:

• Warmup and training steps passed to get_scheduler are scaled by accelerator.num_processes.
• Step counts used to build the schedule account for the sharded dataloader length (math.ceil(len(train_dataloader) / accelerator.num_processes)).
• After accelerator.prepare, a warning is logged if the prepared dataloader length deviates from that expectation.
• Added test_t2i_adapter_sdxl_num_train_epochs in examples/t2i_adapter/test_t2i_adapter.py to cover the --num_train_epochs path.

Coordination

Covered by the ongoing community tracker #8384. cc @sayakpaul @geniuspatrick

Minimal training command using num_train_epochs

export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
export OUTPUT_DIR="sdxl-t2i-adapter-model"

accelerate launch train_t2i_adapter_sdxl.py \
  --pretrained_model_name_or_path=$MODEL_DIR \
  --output_dir=$OUTPUT_DIR \
  --dataset_name=fusing/fill50k \
  --mixed_precision="fp16" \
  --resolution=1024 \
  --learning_rate=1e-5 \
  --num_train_epochs=2 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --seed=42

Tests I ran

pytest examples/t2i_adapter/test_t2i_adapter.py
python3 -m py_compile examples/t2i_adapter/train_t2i_adapter_sdxl.py examples/t2i_adapter/test_t2i_adapter.py
ruff check examples/t2i_adapter/train_t2i_adapter_sdxl.py examples/t2i_adapter/test_t2i_adapter.py
ruff format --check examples/t2i_adapter/train_t2i_adapter_sdxl.py examples/t2i_adapter/test_t2i_adapter.py

All passed (2 passed in 84.45s).