[python] Add LeRobot Dataset v3 import

[XiaoHongbo-Hope]

Purpose

Add a one-time image-based LeRobot v3 importer:

LeRobot Dataset v3 -> Arrow batches -> Paimon table

snapshot_id = conn.load_from_lerobot(
    "robot_data",
    "/path/to/lerobot_dataset",
)

• Supports local directories, FileIO URIs such as OSS through source_options, and Hugging Face repo_id.
• Derives the schema from meta/info.json, preserves feature names and Episode boundaries, writes Arrow batches, and creates one snapshot.
• Maps scalars to scalar types, vectors to VECTOR, higher-rank tensors to nested ARRAY, and compressed images to BLOB.
• Reuses the HDF5 structural Arrow validation and adds strict LeRobot dtype, shape, component-name, BLOB, and VECTOR compatibility checks for existing tables.
• Keeps LeRobot optional through pypaimon[lerobot] and rejects v2.x clearly.

Remote FileIO sources are read directly without materializing the dataset locally. Episode metadata projects only the five routing columns; the current implementation loads one active data Parquet file into memory before slicing it into batches.

This PR supports image features only. Video features are rejected with a clear error; packed video import will be a separate follow-up after its storage model is finalized. This PR does not depend on packed video support.

Tests

• Official LeRobot API creates a two-Episode v3 image dataset in a temporary directory without network access.
• Covers metadata-derived schema, scalar/vector/tensor values, ordering, image BLOB round trip, append, snapshot, shape/dtype/component-name/BLOB/VECTOR incompatibility, safe numeric conversion, empty datasets, OSS reads, missing dependency, v2 rejection, and video rejection.
• multimodal_lerobot_test.py: 17 passed.
• Flake8, py_compile, and git diff --check: passed.

[JingsongLi]

[P1] Preserve LeRobot component names in the append contract. This description records only dtype and shape, so two valid features such as names=["x", "y"] and names=["y", "x"] produce identical schemas and _validate_lerobot_schema accepts the append. The loader then writes vector positions unchanged, silently mixing different component semantics in one training table. Please persist a canonical representation of feature["names"] in the field metadata and compare it for existing tables; add a regression test with reordered names.

[JingsongLi]

[P1] Project the episode metadata columns before materializing rows. _read_remote_parquet currently reads every column and to_pylist() retains all of them for the full import. Normal LeRobot v3 episode rows include per-feature stats/* arrays, while this class only uses five scalar lookup fields. Large FileIO datasets can therefore consume GBs of Python heap and OOM before the first data batch. Please read only episode_index, dataset_from_index, dataset_to_index, data/chunk_index, and data/file_index.

[JingsongLi]

[P2] Handle an empty local dataset before opening LeRobot. A valid zero-frame directory created by LeRobotDataset.create() has meta/info.json but no data/episode/task Parquet files. Opening it here makes LeRobot treat the missing files as a cache miss and attempt a Hub download using the synthetic local/pypaimon-import repo ID, so the documented None no-op path at line 108 is never reached. Please create or validate the table from local_info, then return None before constructing LeRobotDataset when total_frames == 0.

[JingsongLi]

[P2] Validate source values before coercing them to the target Arrow type. pa.array(values, type=field.type) is not a safe cast: for example, a Parquet float value 1.5 under int32 metadata becomes 1, and an oversized float becomes inf under float32. Because strict_arrow_table runs afterwards, it sees only the already-coerced target type and cannot reject the mismatch. Please validate the raw Arrow element type/range or use an explicitly safe cast before rebuilding the column.

[JingsongLi]

+1