Jaap/fix coc oloader max individuals - #3435

Fixes #3432.

COCOLoader.get_dataset_parameters derived bodyparts/individuals from train.json and ignored model_cfg.metadata, while DLCLoader treated the config as authoritative.

max_individuals had no effect on COCO projects, and training/evaluation could crash deep inside pad_to_length whenever an image had more individuals than the busiest training image.

• model_cfg.metadata is now authoritative for bodyparts/individuals, matching DLCLoader.
• Fail fast: COCOLoader now validates train.json/test.json against model_cfg.metadata at construction time, raising ValueError if an image has more individuals than the model supports, or if the annotated bodyparts don't match the configured ones.
• get_project_parameters now optionally takes test_json and suggests max_individuals across both splits, instead of only train.json. docssting is updated to describe the non-authorative bootstrapping role of the method.
• Updated the README and dev-docs snippets.
• Added tests/pose_estimation_pytorch/data/test_cocoloader.py covering:
  • get_dataset_parameters uses model_cfg capacity (not the JSON) when it covers both splits
  • Construction raises ValueError when an image (train or test) exceeds the configured capacity
  • Construction raises ValueError on bodypart mismatch
  • Loader works without a test.json
  • get_project_parameters suggestions cover both train and test JSON, raise on an empty dataset, and warn on multi-category JSON (both splits)