feat(lerobot_v3): add image-mode output to the v3 writer

[griswaldbrooks]

Adds an image-mode output option to the LeRobot v3 writer: camera frames can be stored as
PNG images inline in the data parquet instead of being encoded to MP4. Video mode remains
the default and its output is unchanged.

What changed

Writer (forge/formats/lerobot_v3/writer.py) — new LeRobotV3WriterConfig.video flag
(default True), following upstream LeRobot's own video=False convention. With
video=False:

• each camera frame is PNG-encoded into the data parquet as a struct<bytes, path> column,
  matching the HF datasets.Image() storage LeRobot uses for image datasets;
• the feature is declared in info.json as dtype: "image";
• the video encoder is never called and no videos/ directory is created;
• video_path and video_files_size_in_mb are null, with the keys retained because v3
  loaders require them to exist.

Pillow is required for image mode (added to the [lerobot] extra) and checked up front.

Reader (forge/formats/lerobot_v3/reader.py) — camera features with dtype: "image" are
decoded from the inline PNG bytes (or a referenced path) through a lazy PIL loader, so
image-mode datasets round-trip through Forge.

Converter (forge/convert/converter.py, forge/config/models.py) — writer_config is
carried through ConversionConfig.to_dict() and applied in the parallel worker, so
--workers > 1 honours image mode. The sequential and parallel paths now share one
_configure_writer helper.

Usage:

fps: 10
robot_type: pusht
writer_config:
  video: false

Tests

• tests/test_lerobot_v3_writer.py — image-mode structure (info.json dtype, no videos/,
  frame counts, null video_path), plus a video-mode-unchanged regression class.
• tests/test_lerobot_v3_roundtrip.py — image-mode frames survive write → read through the
  real v3 reader.
• tests/test_converter.py — writer_config round-trips through to_dict/from_dict, and
  the parallel worker actually applies video=False.
• tests/test_lerobot_v3_upstream_validity.py — extended with a written_v3_image fixture and
  a TestUpstreamLoaderImageMode class, gated on lerobot being installed like the existing
  upstream-validity tests.

Verification against stock LeRobot

lerobot is not a test dependency, so the spec check was run manually. In a scratch venv with
lerobot 0.4.4, the real 206-episode pusht zarr sample was converted through
forge convert --config in image mode and loaded with stock LeRobotDataset:

lerobot metadata: robot_type=pusht fps=10 episodes=206 frames=25650
LeRobotDataset len: 25650
item 120: observation.images.img tensor (3, 96, 96) dtype=torch.float32 range=[0.255,1.000]

image_keys is populated and video_keys is empty. The gated upstream-validity suite passes
23/23 against real lerobot. Frames decoded by Forge's reader, by stock HuggingFace datasets
via Image(), and by LeRobotDataset are all pixel-identical to the source zarr.

Video mode was compared base vs branch on the same sample: every parquet and JSON artifact is
byte-identical. The MP4 files differ byte-wise, but decoded pixels match, and two runs of the
same build differ by the same margin — x264 run-to-run nondeterminism, not a behaviour change.

Known limitations

Filed as follow-ups, all pre-existing or narrow enough to ride a patch release:

• Parallel conversion writes every episode to chunk-000 #1 — parallel conversion writes every episode to chunk-000 (pre-existing; affects video
  mode equally).
• LeRobot v3 writer: non-uint8 frames clip in image mode but wrap in video mode #2 — non-uint8 frames clip in image mode but wrap in video mode.
• LeRobot v3 writer: 256-row sampling window can misdeclare a late-appearing image column #3 — the 256-row sampling window in finalize() can misdeclare a late-appearing image column.
• LeRobot v3 writer: a camera absent for a whole episode is dropped from that episode's parquet #4 — a camera absent for a whole episode is dropped from that episode's parquet while
  info.json still declares it.
• LeRobot v3 writer: null feature cells now surface at load time instead of convert time #5 — null feature cells now surface at load time rather than convert time.