Fix GPU selection and execution path handling

[Traveler-Roy]

• Fix GPU selection defaults and priority:
  • default gpu_id now falls back to 0
  • explicit CLI --gpu overrides CUDA_VISIBLE_DEVICES
  • omitted --gpu preserves existing CUDA_VISIBLE_DEVICES
• Add optional --gpu support across main.py, run_pipeline.py, and prepare_data.py.
• Propagate explicit GPU selection from parent pipeline commands to child processes.
• Update shell wrappers so they only forward GPU selection when the user explicitly provides it.
• Harden DDP/rank-zero artifact handling for logging, evaluation, visualization, LoRA adapters, and BOW outputs.