Fix BatchSamplerShard tail sampling

[taking-lying-flat]

Summary

• preserve the full dataset size when constructing BatchSamplerShard
• generate the global sample order before applying distributed sharding
• truncate for drop_last=True and repeat-pad for drop_last=False
• apply the same distributed semantics to shuffled, sequential, and length-grouped sampling
• compute batch sampler length from the per-rank sample count

Root cause

BatchSamplerShard previously floored the dataset size by the DP world size in its constructor. The non-length-grouped paths then generated indices only for that shortened prefix, so tail samples were permanently excluded even when drop_last=False. For packed datasets, the omitted item could be a full packed bin.

Impact

All original samples now participate when drop_last=False, while every DP rank still receives the same number of samples. When drop_last=True, indices are truncated after the epoch-specific global order is generated.