Add mixture-of-experts latent dynamics and stronger dynamics experiment bundle

[abelobsenz]

Motivation

• Improve surface-forecasting performance vs the tree baseline by modeling heterogeneous latent transition regimes that a single deterministic transition may miss.
• Encourage stronger long-horizon signal learning and robustness by enabling multi-step rollouts with scheduled sampling and an MoE transition head.

Description

• Implemented a configurable mixture-of-experts latent dynamics path inside IVDynamicsTorchModel by adding ModelConfig.dynamics_n_experts and emitting a dynamics head of size latent_dim * n_experts when n_experts > 1.
• Added a small context/asset-conditioned gating network and softmax blending of expert latent deltas in forward_dynamics, keeping behavior identical when n_experts == 1.
• Exposed dynamics_n_experts through the training plumbing via TrainingConfig and the CLI (--dynamics-n-experts), and persisted dynamics_n_experts in the training summary metadata.
• Strengthened the dynamics_bundle experiment recipe in the plan to push longer-horizon learning by setting dynamics_n_experts=4, multi-step rollouts (rollout_steps=3), rollout_random_horizon=True, scheduled sampling (rollout_teacher_forcing_start/end), and larger rollout_surface_lambda.

Testing

• Ran python -m compileall ivdyn successfully to ensure the modified modules compile without syntax errors.
• Verified CLI wiring and option exposure with python -m ivdyn train --help | rg -n "dynamics-n-experts|rollout-teacher-forcing" and python -m ivdyn experiment-plan --help | rg -n "dynamics-n-experts", which confirmed the new flags and bundle settings are present.
• Performed quick smoke checks of python -m ivdyn --help to ensure the package CLI remained functional.
• Full training/evaluation benchmark runs against dataset.npz were not executed because no local dataset file was available in the environment, so empirical improvement vs the tree baseline should be validated by running the updated dynamics_bundle on a representative dataset (recommended: run python -m ivdyn experiment-plan --execute or python -m ivdyn train ... --dynamics-n-experts 4 with your dataset).

---

Codex Task