test(mot): add comprehensive boundary tests for MoT module in training/eval modes (#54)

[9-71]

Summary of Changes

Addresses #54 by expanding boundary test coverage for the MoT (Mixture-of-Tracking) module under edge-case spatial dimensions and routing configurations, ensuring both forward pass integrity and backward gradient propagation during training.

Key Additions to tests/test_mot.py:

1. Window Size > Feature Map (Boundary a):

   • Added test_mot_window_expert_handles_window_larger_than_feature_map_train and test_mot_block_handles_window_larger_than_feature_map_train.
   • Verifies padding/cropping logic and full gradient flow to experts & router when window_size exceeds input (H, W).

2. Odd Spatial Dimensions Shift (Boundary b):

   • Added test_mot_window_expert_shift_odd_spatial_train and test_mot_block_shift_odd_spatial_train.
   • Ensures shift/unshift operations remain stable on odd grid sizes (e.g., 7x9, 9x11) in training mode.

3. Exploration Epsilon Behavior (Boundary c):

   • Added test_mot_exploration_eps_active_in_train_disabled_in_eval and test_mot_exploration_eps_strict_zero_in_eval.
   • Explicitly validates dense routing during train() vs strictly sparse routing (Top-1) during eval().

---

Test Results

• Unit Tests: 33/33 passed (pytest tests/test_mot.py).
• Code Stability: Confirmed that ultralytics/nn/modules/mot/ handles all edge cases gracefully without requiring invasive runtime patches.
• Lint Status: All new test code is compliant with ruff.

[isLinXu]

LGTM