fix(sft): expose filtered row reasons

[lsr-05]

Summary

SFT now reports why rows were filtered when no trainable rows remain.

• count empty prompts/responses, all-prompt rows, and prompt/response budget violations separately
• include the counts in the existing ValueError and filter log
• cover empty datasets and mixed all-invalid datasets in the CPU suite
• keep _record_to_train_sequence behavior unchanged for existing callers

This is a follow-up to #90.

Validation

• python -m pytest -q tests/test_trainer_dataset_utils_cpu.py tests/test_qwen35_multimodal_cpu.py (26 passed, 11 skipped)
• ruff check areno/api/trainers/sft.py tests/test_trainer_dataset_utils_cpu.py

[lsr-05]

#79