- Add SeqlenBalancedSampler based on Karmarkar-Karp algorithm to balance
  sequence lengths across DP ranks for GRPO training
- Add streaming mode support for StreamingDataset via
  should_check_consumption_status parameter
- Add polling_mode sampler cache lookup in controller to avoid redundant
  sampling when data is insufficient
- Replace print() with logger.info() in controller
- Downgrade 1D tensor warnings to info level in client and metadata
- Add comprehensive unit tests for SeqlenBalancedSampler and KarmarkarKarp

Signed-off-by: 宁本哲 <ningbenzhe@xiaohongshu.com>

Signed-off-by: 宁本哲 <ningbenzhe@xiaohongshu.com>

Signed-off-by: 宁本哲 <ningbenzhe@xiaohongshu.com>