[GenEval] feat: add multi-gpu in evaluations & tqdm display the progress

[lxazjk]

Summary

geneval/evaluation/evaluate_images.py previously evaluated images sequentially on a single GPU, which is slow for large-scale runs and provides limited visibility into progress.
This PR adds a simple, low-intrusion multi-GPU sharding mode to improve throughput and adds progress bars for better UX.

What’s changed

Multi-GPU evaluation via multiprocessing sharding

Collects all (image_path, metadata) tasks first, then splits them across GPUs and runs one process per GPU.
Each worker sets its CUDA device, loads models locally, evaluates its shard, and returns results to the main process for aggregation.

Progress reporting with tqdm Single-GPU mode shows a single overall progress bar.

Multi-GPU mode shows per-GPU worker progress bars.
Stability improvements for spawn + CLIP color classification
Refactors the crop dataset used by CLIP color classification to avoid relying on global variables in DataLoader worker processes.
Adds clip_num_workers option and defaults to 0 when num_gpus > 1 to avoid nested multiprocessing issues.

How to run

Single GPU

python geneval/evaluation/evaluate_images.py <imagedir> --outfile results.jsonl --model-path <ckpt_dir>                                                                                     

Multi GPU

python geneval/evaluation/evaluate_images.py <imagedir> --outfile results.jsonl --model-path <ckpt_dir> --options num_gpus=8 gpu_ids=0,1,2,3,4,5,6,7 clip_num_workers=0