fix(moe): preserve pruned experts' kernel sizes in model.yaml so retraining reloads their weights

[Lfan-ke]

Problem

MoEPruner keeps the highest-utilization experts, which in a trained model have heterogeneous depthwise kernel sizes (e.g. a layer keeps experts with kernels [5, 9]). PR #192 encoded the reduced num_experts into model.yaml, but not the kept experts' kernel sizes.

Consequently YOLO(pruned.pt).train() (the prune -> LoRA / full fine-tune recovery workflow) rebuilds each ES_MOE from YAML with the default kernels [3, 5, ...][:num_experts], which mismatch the kept experts' actual kernels. intersect_dicts then drops those depthwise weights on the shape mismatch and re-initializes the experts randomly, so retraining starts from a degraded model.

On a VisDrone YOLO-Master-EsMoE-N pruned to 2 experts/layer this is reproducible: direct-inference val is mAP50 0.4269, but the first retrain epoch starts at mAP50 ~0.002 (random experts). After the fix the same retrain starts from a correctly loaded model.

Fix

ES_MOE gains an optional expert_kernel_sizes argument (defaults to None, so behavior is unchanged when unset). MoEPruner._sync_yaml_num_experts now writes the full ES_MOE positional arg list, including each kept expert's kernel size, into model.yaml. The YAML rebuild then reconstructs the exact kept experts and their weights survive intersect_dicts.

Tests

tests/test_moe_prune_yaml_sync.py gains test_sync_preserves_expert_kernels, which builds a stand-in pruned model with non-default kernels, syncs, rebuilds, and asserts the kernels and weights are preserved. All three tests pass. The change is backward compatible.

[Lfan-ke]

The three failing checks (Tests 3.13 ubuntu/windows, Tests 3.8 torch1.8) are pre-existing on main at the same base commit d3341f3, not introduced here. The CI run for main fails the identical three jobs: https://github.com/Tencent/YOLO-Master/actions/runs/30513506266 .

This change is isolated to MoE pruning YAML sync (ES_MOE.expert_kernel_sizes + MoEPruner._sync_yaml_num_experts), and its regression tests in tests/test_moe_prune_yaml_sync.py pass locally. The 3.8 torch1.8 failure is the legacy-autocast issue already addressed by #185. The 3.13 failures are in the unrelated MoT/DDP test suite.

[isLinXu]

LGTM