(draft) feat: add LeWM, an action-conditioned latent world model trained with SIGReg

[gabrielfruet]

LeWM is a latent world model: it predicts the next frame's embedding from past embeddings and the action taken, never reconstructing pixels. SIGReg, already here for LeJEPA, prevents collapse, so there is no teacher, no EMA and no stop-gradient, and the encoder trains from pixels alongside the predictor. Paper: https://arxiv.org/abs/2603.19312

• lightly/models/modules/world_model/: LatentDynamicsPredictor, a causal transformer with AdaLN-Zero action conditioning and a rollout() that feeds predictions back over a sliding window, plus an ActionEncoder MLP.
• lightly/loss/: LeWMLoss (prediction MSE plus lambda_param * SIGReg, default 0.1) and latent_distance, a new l1/l2 helper with optional layer norm for later methods.
• A PyTorch example, docs, and ~500 lines of tests.
• scaled_dot_product_attention goes through getattr, with a manual fallback, so type checking passes on torch 1.10.

The API covers only what LeWM needs, the first of a planned sequence; later methods add keyword arguments defaulting to this behavior. Missing: the Lightning example variants and the README table row.

The example synthesizes its own trajectories, a square pushed by the action, so it needs no simulator. Loss falls from 1.05 to 0.33 over 10 epochs. That shows the loop trains, not that it reproduces the paper. 12 CI checks green.