Dynamic LLM Trainer Framework for Kubeflow — TRL Backend with Pluggable Multi-Framework Support

Kubeflow Trainer V2 currently supports only TorchTune as its LLM fine-tuning backend. TorchTune stopped adding new features in July 2025, leaving Kubeflow with ~4 supported models and no support for DPO, GRPO, or other post-training methods. This project builds a pluggable multi-backend framework for Kubeflow Trainer. The first new backend is TRL (HuggingFace's fine-tuning library), adding support for SFT, DPO, GRPO, Unsloth acceleration, and multi-node DeepSpeed. The architecture includes: an LLMBackend ABC + registry in the Python SDK for config-driven backend dispatch, a LLMFrameworkHandler interface in the Go control plane (Torch plugin) for zero-touch extensibility, and entry_points-based registration so the community can add backends like LlamaFactory without upstream changes. Deliverables: TRL backend (Python SDK + Go handler), pluggable registry, runtime configs, integration tests, and documentation. All existing TorchTune workflows remain unchanged.