[python] Support custom dataset readers in PaimonLeRobotDataset

[XiaoHongbo-Hope]

Purpose

Let PaimonLeRobotDataset read logical frames through custom dataset readers.

Changes

• Add PaimonDatasetReader: metadata, logical frame schema, validated batched row reads, delta windows, media decoding, and Torch conversion.
• Make PaimonLeRobotDataset a Dataset facade over PaimonDatasetReader.
• Accept a standard Paimon table or a configured PaimonDatasetReader through the same constructor.

Custom readers implement read_indices. schema enables eager validation; file_io is required only for BLOB or video descriptors. Reader instances must be picklable for DataLoader workers.

Relationship to LeRobot

LeRobot main selects a BaseDatasetReader by meta.storage_format. The reader owns row access, delta windows, media decoding, and returns assembled samples.

The BaseDatasetReader source docstring explicitly lets subclasses define their own constructors because storage inputs differ. schema and file_io are PyPaimon-specific inputs, not LeRobot parameters.

# LeRobot main
LeRobotDataset
└─ BaseDatasetReader
   ├─ DatasetReader          (Parquet / MP4)
   └─ LanceDatasetReader     (Lance)

# This PR
PaimonLeRobotDataset
└─ PaimonDatasetReader
   ├─ _PaimonTableDatasetReader   (standard table group)
   └─ CustomDatasetReader         (custom storage layout)

# Future LeRobot backend
LeRobotDataset
└─ PaimonDatasetReader       (BaseDatasetReader)

PaimonDatasetReader mirrors the official Reader responsibilities. Once the official contract is available in PyPaimon's supported LeRobot range, it can implement BaseDatasetReader without changing custom row-resolution code.

Official references: BaseDatasetReader, storage_format reader lookup.

Review focus

Is the minimum custom-reader contract (read_indices, plus optional schema and file_io) sufficient while keeping sample assembly in one Reader?

Tests

• 80 passed in multimodal_lerobot_test.py.
• Python syntax compilation, flake8 for changed files, and git diff --check pass.

[JingsongLi]

Reviewed 1cf87f1. Requirement fit: SUPPORTED. Implementation: CLEAN in the reviewed reader contracts.

Custom storage layouts can now provide logical frames while reusing the existing episode selection, delta windows and sample assembly. The boundary is small enough to be useful: read_indices supplies rows, and the shared reader validates completeness, uniqueness and control columns before assembling samples. The facade also preserves the existing table-backed path and explicitly avoids forwarding pickle special methods. The LeRobot reader reference supports this division of responsibilities; direct integration with LeRobot's backend registry remains future work.

I checked batch ordering, episode remapping, schema/media contracts and serialization ownership. The available environment passed 39 tests and additional real-Arrow custom-reader probes; 38 optional tests were skipped. Three full-file failures reproduce unchanged on the exact base. Torch/media/DataLoader end-to-end execution was unavailable locally; current PR CI is successful. No actionable introduced regression found.