feat: enhance SFT training process with new tokenization method, implement custom trainer class for loss computation, and update README with GRPO launcher details for Unsloth LoRA integration

fix: update README with SFT training configuration details, modify modal training scripts to disable assistant-only loss and packing for compatibility, and adjust test assertions to reflect these changes

feat: expand README with synthetic SFT dataset generation instructions, enhance dataset verification and pushing to Hugging Face Hub, and improve modal training scripts with default configurations for curriculum and GPU fallback

feat: introduce reward ablation configurations for enhanced training flexibility, implement YAML loading with extends support, and add reward variant tracking in training scripts