fix(moe): keep pruned expert count in model.yaml so pruned models survive retraining

[Lfan-ke]

Problem

MoEPruner physically reduces each ES_MOE's expert list and router, but the saved model.yaml still describes the original expert count. When a pruned checkpoint is then used for further training (the prune -> LoRA / full fine-tune recovery workflow), Trainer.setup_model rebuilds the model from model.yaml with ES_MOE's default expert count, and BaseModel.load()'s intersect_dicts silently drops the reduced expert/router weights on shape mismatch. The model re-inflates to the default expert count with randomly initialized experts, so "recovery" ends up training a reconstructed non-pruned model rather than the pruned one.

Fix

After surgery, MoEPruner writes each pruned ES_MOE's expert count into its model.yaml args ([out_channels, num_experts]; ES_MOE clamps top_k = min(top_k, num_experts)). The pruned architecture now round-trips through a YAML rebuild, so retraining/fine-tuning operates on the actual pruned model.

Verification

Added tests/test_moe_prune_yaml_sync.py: reproduces the re-inflation when the yaml is not synced, and confirms the sync preserves the pruned expert count with all weights round-tripping.

[isLinXu]

LGTM