fix(data): fill sample requests across multiple epochs

[feng1201]

When a request spans more than one dataset boundary, RolloutDataSource.get_samples only wraps once. With three prompts, get_samples(8) returns six groups and stores sample_offset=5; subsequent reads also diverge from the dataset stream.

Read across as many epochs as needed, preserving epoch shuffling, group/sample indices, buffered samples, and checkpoint resume. Reject positive requests from an empty dataset so the loop cannot stall. The regression uses actual data-source classes, a JSONL file, and a locally saved Transformers tokenizer; no model download or mocked data-source implementation is needed. The test is included in the existing CPU CI matrix.

Validation:

• Original source: 12 failed, 13 passed in the new regression module.
• Fixed source: python -m pytest tests/test_data_source.py tests/test_read_file_slicing.py tests/test_filter_long_prompt.py tests/test_sample.py -q — 52 passed.
• All applicable pre-commit hooks passed on the four changed files.
• Python 3.11.15, PyTorch 2.13.0 CPU, Transformers 5.15.0, Ray 2.55.0 on macOS arm64. Full GPU training was not run.

AI disclosure: implementation and local verification were performed with OpenAI Codex.