[vlm, data] feat: expose Energon read-order controls - #139

Summary

• expose --data-shuffle-buffer-size and --data-max-samples-per-sequence for multimodal training
• forward the resolved controls to both direct and blended Energon dataset paths
• report the active read-order settings on rank zero while preserving existing defaults
• add regression coverage for value resolution, CLI parsing, and both loader branches

Motivation

Offline-packed WebDataset records can contain long runs of similarly sized packs. These controls let users increase sample-level randomization and create shorter shard sequences without changing the default data pipeline.

Validation

• uv run --no-project --isolated --with pytest pytest -q tests/test_vlm_dataloader_read_order.py — 6 passed
• python3 -m py_compile loongforge/data/multimodal/dataloader_provider.py loongforge/train/arguments.py tests/test_vlm_dataloader_read_order.py
• python -m build --sdist --wheel
• twine check on the generated source and wheel distributions
• pre-commit run spdx-check on all changed Python files
• git diff --check