feat(moe): add validation-driven dynamic top-k annealing scheduler

[Lfan-ke]

What

Adds DynamicTopKScheduler: a validation-driven scheduler that anneals the number of active experts (top_k) during training, instead of the auxiliary balance-loss coefficient that existing MoE schedulers tune.

Each validation epoch it drops top_k by one when validation mAP has saturated (sliding-window) and expert usage is balanced (EMA Gini ≤ ceiling), and rolls back + freezes if the step hurts mAP. A converged, balanced model then runs with fewer active experts — lower inference FLOPs, matching the ES-MoE "compute-on-demand" premise, and complementary to post-hoc MoEPruner.

API

• DynamicTopKScheduler / TopKAnnealConfig / apply_top_k(model, k) in ultralytics/nn/modules/moe/topk_anneal.py.
• Opt-in training callback: model.add_callback("on_fit_epoch_end", create_topk_anneal_callback(cfg)).
• Fully backward compatible: disabled by default (enabled=False); the callback is defensive and never interrupts training if MoE modules or optional deps are absent.
• Reuses mean_usage_gini_from_model and mirrors the apply_balance_loss_coeff pattern.

Tests

tests/test_moe_topk_anneal.py — 9 tests (anneal on saturation+balance, no-anneal when imbalanced / still improving, rollback on mAP drop, floor at min_k, state_dict roundtrip, disabled no-op, callback never raises, val-map extraction). All pass.

Closes #52