[KD][MoE] Add a validated MoE knowledge distillation recipe / example under examples/llm_kd/

Summary

MoE fine-tuning is well supported in AutoModel and validated at scale (GLM-5, MiniMax-M2.5, DeepSeek-V3.2, Step-3.5), but MoE knowledge distillation has no counterpart, there is no MoE example under examples/llm_kd/, and the KD recipe has diverged from train_ft.py in ways that break EP paths that already exist elsewhere in the codebase.

This request is for a validated, end-to-end MoE KD example covering the common combinations users are trying today. Even after the individual bugs are fixed, a working reference config is what turns "possible in principle" into "someone can copy-paste and run."

Ask
A recipe example (or set of examples) under examples/llm_kd/ covering at minimum:

1. Dense student + MoE teacher using separate_meshes (once released). Reference pair: something small enough to fit on a single node Qwen3-4B student ← Qwen3-30B-A3B teacher is realistic on 8×H200.
2. MoE ↔ MoE Reference pair: Qwen3-30B-A3B student ← larger MoE teacher, or smaller MoE ← Qwen3-30B-A3B teacher.
3. Documented ep_size guidance for the student and teacher meshes, with a note on the ep_shard_size == 1 interaction with offload