Add HuggingFace Hub integration for dataset/model publishing

[Edmon02]

Description

Add first-class integration with HuggingFace Hub to publish generated trajectories as a dataset and push trained models.

Tasks

• TrajectorySerializer.push_to_hub() — upload JSONL trajectories as a HF dataset
• Dataset card generation with statistics (trajectory count, language distribution, reward histogram)
• Model card generation after training (VRAM budget, base model, LoRA config, benchmark results)
• trainer.push_to_hub() — merge LoRA weights and push to HF Hub
• Notebook cells for one-click publishing

Benefits

• Share generated datasets with the community
• Easily load pre-generated trajectories: datasets.load_dataset("Edmon02/rfa-trajectories")
• Push trained models for inference: transformers.AutoModelForCausalLM.from_pretrained("Edmon02/rfa-coder-3b")