test(ddpg): deeper DDPG coverage — multi-episode critic-loss trend + in-loop Polyak soft-update

[belarusian]

TICKET-054 / issue #122.

Adds two closed-loop DDPG tests to tests/test_ddpg_integration.py (no production-code change):

• test_multi_episode_critic_loss_trend: runs a deterministic closed-loop, multi-episode training loop (8 episodes x 12 steps, critic update per step through the public get_allocation API) and asserts the critic actually learns — the mean critic loss over the later third of updates is at least 25% below the earlier third (measured ~57-68% reduction for seed=42).
• test_soft_update_targets_track_online_in_loop: verifies the Polyak soft-update (_soft_update_targets, the real update used by the production loop in core.py) in context — after training, target actor/critic weights are close to but not identical to the online weights, and the target critic's Q-estimate on a probe batch is close to the online critic's.

Closes #122