feat(data_parsing): L2D 1 Hz sequential multi-view windows for the World Model (#16)

[gcordova10]

What this adds

L2DDataset currently loads only the current frame and leaves visual_history
as zeros, with no future frames — so the World Model / JEPA (#13, #85) has neither
history nor targets. This PR adds sequential 1 Hz multi-view windows (the
"Sequential frame access (needed for #13)" checklist item in #16).

Design (testable without lerobot)

• data_parsing/l2d/world_model_windows.py (new, pure / dataset-agnostic):
  • stride_for_hz(source_hz, wm_hz) — L2D 10 Hz → 1 Hz = stride 10.
  • window_offsets(N, stride) — past [-(N-1)s..0] (oldest→newest, current last)
    and future [s..N*s].
  • required_margins(N, stride) = ((N-1)*s, N*s).
  • build_windows(load_frame, row, ep_start, ep_end, num_frames=4, stride=10) →
    (history_frames, future_frames), each [N, V, 3, H, W]; raises IndexError
    if a window would cross the episode boundary. Takes a load_frame(row)->[V,3,H,W]
    callable, so it has no dataset/lerobot dependency and is unit-tested with a
    synthetic loader.
• data_parsing/l2d/dataset.py (opt-in, default OFF → byte-identical to before):
  • new args include_world_model_windows=False, wm_num_frames=4, wm_hz=1.0, source_hz=10.0;
  • _load_multiview_frame(row) refactor, reused for the current frame and every
    window frame;
  • _build_sample_index margins take the max of the egomotion window (64/64)
    and the World Model windows, so a valid frame always has a complete window;
  • __getitem__ emits history_frames / future_frames when enabled;
    L2DSample gains the two optional keys (NotRequired, version-guarded import
    for py3.10/3.12).

How it fits

• Proposal: Feature Reconstruction Loss for FutureState (Auxiliary Self-Supervised Task) #13 train_il: compute_step_loss consumes exactly batch["history_frames"]
  / ["future_frames"] → encode_history → predict_future → jepa_loss (these were
  synthetic before).
• feat(world-model): slow World Model branch (JEPA) + AutoE2E wiring (#13) #85 World Model: history_frames feeds WorldActionModel; future_frames
  are the JEPA targets.
• Does not overlap Proposal: World Action Model (1 Hz camera multi-view encoder without BEV view + encoded features temporal fusion) #93 (different area — data_parsing).

1 Hz / 10 Hz

L2D is 10 Hz (egomotion.py, _DT=0.1), so source_hz=10.0 (stride 10).
Parametrized: a 30 FPS source → source_hz=30.

Scope / not in this PR (kept honest)

• Camera calibration (BEV camera_params): TODO: Dataset Selection and DataLoader Implementation #16 lists it, but I keep this PR to
  the sequential windows. l2d/camera.py already has make_camera_params_placeholder()
  (identity) with a documented TODO to parse intrinsic @ extrinsic → [3,4] per view
  from L2D's extrinsic_RDF.yaml; BEV fusion runs on its learnable pseudo-projection
  meanwhile (not blocking). Follow-up: the real YAML parser.
• Performance: decoding N×2 multi-view frames per sample is expensive; for real
  training prefer pre-extraction (data_parsing/pre_extracted.py) or caching. The
  functional correctness here is independent.

Tests

pytest Model/tests/test_world_model_windows.py → 14 passed. ruff + mypy clean.
Diff: 2 new files + 2 edited (dataset.py, __init__.py).

[gcordova10]

CI green . This implements the "Sequential frame access (needed for #13)" item from #16: 1 Hz multi-view past/future windows for the World Model.

• New pure, dataset-agnostic module data_parsing/l2d/world_model_windows.py (stride_for_hz, window_offsets, required_margins, build_windows) — unit-tested without lerobot (14 tests) by injecting a synthetic frame loader.
• L2DDataset extended opt-in (include_world_model_windows=False by default → unchanged otherwise): emits history_frames / future_frames [N, V, 3, H, W], with episode-boundary checks.

Feeds the JEPA directly: #13's compute_step_loss consumes batch["history_frames"] / ["future_frames"], and #85's World Model uses them as input / targets (they were synthetic placeholders before). Independent of #85 (different area) and doesn't overlap #93.

Kept honest about scope (details in the PR body): BEV camera calibration is left as a documented TODO in l2d/camera.py (BEV runs on the learnable pseudo-projection meanwhile), and pre-extraction/caching for decode throughput is a noted follow-up. Ready for review.

[m-zain-khawaja]

approved - thanks @gcordova10