Detect TRL trainer class: route DPO/ORPO, mark GRPO cross-cuts

Add detect_trainer_class() which scans code cells for TRL trainer
class names (GRPOTrainer, DPOTrainer, ORPOTrainer, KTOTrainer,
RewardTrainer, PPOTrainer) and returns the highest-priority match.
This is the authoritative signal for what kind of RL/preference
training a notebook runs, regardless of the filename.

Three behavior changes driven by the trainer class:

1. DPO / ORPO routing to the GRPO section
   Zephyr (7B) DPO and Llama3 (8B) ORPO were landing in the Mistral
   and Llama sections because they follow the filename-based
   architecture routing. They are now force-routed to the GRPO &
   Reinforcement Learning section with Type set to "DPO" / "ORPO"
   respectively.

2. (GRPO RL) suffix on cross-cut sections
   A notebook that trains with GRPOTrainer may also appear in
   cross-cutting sections (Vision (Multimodal), Embedding, OCR). In
   those non-GRPO sections the Type now gets a "(GRPO RL)" suffix so
   readers can tell it is an RL notebook and not a plain SFT one.
   Examples in the Vision (Multimodal) section:
     * Qwen2.5 VL (7B)  Vision Math + vLLM (GRPO RL)
     * Qwen3 VL (8B)    Vision Math (GRPO RL)
     * Qwen3.5 (4B)     Vision Math (GRPO RL)
     * Gemma3 (4B)      Vision Math (GRPO RL)

3. Row rendering moved into the per-section loop
   Previously the row string was built once and appended to every
   section the notebook belonged to. The row is now built inside the
   `for section_name in data["sections"]` loop so the Type column can
   vary per section (to append "(GRPO RL)" only in cross-cut sections).

The vLLM suffix is now only appended for GRPO-class training so the
new DPO / ORPO rows stay clean.

Author: danielhanchen