fix(data): preserve empty datasets during prompt length filtering

[GoldArowana]

An empty JSONL file or a valid @[0:0] slice currently crashes during Dataset(..., max_length=8) construction because filter_long_prompt indexes origin_samples[0]. Return the original empty list before inspecting the first prompt.

This preserves existing Dataset/filter semantics: empty input is already accepted with max_length=None, and nonempty input may become empty after length filtering. It does not establish that empty data can complete rollout, evaluation, or training; downstream consumers retain their own validation.

Extend the existing registered CPU test with actual temporary JSONL files and a locally constructed Transformers fast tokenizer backed by WordLevel (no model downloads). Cover empty files and empty slices with and without a length limit, an 8-token boundary and 9-token overlong prompt, all-filtered input, empty-list identity with and without a processor, and existing ordering behavior. The file retains NUM_GPUS = 0.

Validation on Linux / Python 3.12:

• Before the guard: python -m pytest tests/test_filter_long_prompt.py tests/test_read_file_slicing.py — 4 failed, 20 passed; exit 1, with the expected empty-input IndexErrors.
• After the guard: the same command — 24 passed; exit 0.
• Official CI-shaped entry: python tests/test_filter_long_prompt.py — 14 passed; exit 0, including after formatting.
• Normal commit ran all applicable repository pre-commit hooks successfully; git diff --check passed.
• Independent read-only review found no blocking correctness, coverage, or CPU CI issues.

Tests used CPU-only, offline settings: CUDA_VISIBLE_DEVICES='' SLIME_ACCELERATOR=cpu HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 PYTHONDONTWRITEBYTECODE=1 PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 PYTEST_ADDOPTS='-p no:cacheprovider' PYTHONPATH=.. No GPU, training, rollout service, or weight-download checks were run. This change covers a CPU data-filtering boundary.

Related PRs #2365, #2449, #2126 and #2363 and their discussions were refreshed before publication; their epoch-sampling and multimodal changes do not address this empty-input guard.

AI assistance was used for reproduction, implementation, tests, and an independent review; the reported validation commands were actually executed.