Refactor dataset part to use the new dataset and processor interface

[RayenTian]

Refactor the SFT and GRPO entry script to adopt the new dataset processor interface introduced in RawDataset and the PROCESSOR_REGISTRY, removing ad‑hoc preprocessing and aligning with the GRPO data path.

Motivation

We recently introduced a unified processor interface and registration mechanism for datasets in PR !1506. The SFT path still wires processors and preprocessing via partials in run_sft.py, which diverges from the new interface and duplicates logic. Unifying both paths reduces maintenance, enables consistent configuration, and simplifies documentation.

Scope

1. Replace ad‑hoc processor wiring in run_sft.py with the standard processor interface (TaskDataSpec + PROCESSOR_REGISTRY).
2. Ensure datasets used by SFT can declare/select their processor consistently via config (e.g., data.processor).
3. Make STF dataset inherit RawDataset class. Unified dataset interface.
4. Keep parity with existing SFT features (optional BOS/EOS, add_generation_prompt, optional image preprocessing for CLEVR, etc.) by moving them into the processor layer where appropriate.
5. Move customized OmegaConf op into a common place. Refer to this comment.
6. Unify dataset initialize with super.init. Refer to this comment.