fix blockmask offset compute

[starcrown001]

• Fixed the precision misalignment issue of blockmask when the sequence length (seqlen) exceeds 16k, ensuring correct computation for large-scale inputs.
• This version optimizes the blockmask implementation. Compared to the original (mit-han-lab/Block-Sparse-Attention), achieves an 81% to 188% improvement in forward performance
  and achieves a 48% to 105% improvement in backward performance on H800 Significantly boosts overall operator efficiency.
• Comprehensive regression testing for both accuracy and performance has been conducted against the original flashmask operator. The impact on accuracy or performance is negligible, ensuring compatibility and stability.