Optimize adapter element counting on GPU.

[trivialfis]

• Implement a simple IterSpan for passing iterators with size.
• Use shared memory for column size counts.
• Use one thread for each sample in row count to reduce atomic operations.

The first two items are from #9194 .

The time used for row counts after and before the PR:

 Time (%)  Total Time (ns)  Instances    Avg (ns)      Med (ns)     Min (ns)    Max (ns)   StdDev (ns)                                                  Name                                                
      2.1       44,654,359         32   1,395,448.7   1,393,922.5   1,329,936   1,466,870     61,552.5  void dh::LaunchNKernel<unsigned long xgboost::data::GetRowCounts<xgboost::data::CupyAdapterBatch>(T…
      5.1      121,245,208         32   3,788,912.8   3,793,333.0   3,609,087   3,896,713     94,741.1  void dh::LaunchNKernel<unsigned long xgboost::data::GetRowCounts<xgboost::data::CupyAdapterBatch>(T…