[FEA] Improve the sub-partitioning algorithm for large/skewed hash joins

[firestarman]

PR #7794 has implemented the initial version of joins by the sub-partitioning algorithm for #7471, however there are still some short/long term follow-ups to make it better.

This issue is filed to track the follow-ups.

• [BUG] Two tables with the same key may be placed at different partitions by Table.hashPartition. rapidsai/cudf#13000
  Fixed by Support hashing different types for sub-partitioning #7937
• Support re-partitioning large data multiple times and each time with a different seed. #7911
  (See Implement sub partitioning for large/skewed hash joins #7794 (comment))
  Fixed by Sub-partitioning supports repartitioning the input data multiple times #7996
• Support some special cases to deal with skewed data.
  (See Implement sub partitioning for large/skewed hash joins #7794 (comment))
  Fixed by Sub-partitioning supports repartitioning the input data multiple times #7996
• Make the read batches spillable if they are on GPU in BatchTypeSizeAwareIterator. #7886
• Investigate the perf regression found in NDS runs introduced by Implement sub partitioning for large/skewed hash joins #7794 for non sub-partitioning path.
• [BUG] GPU's output differs from CPU's for big decimals when joining by sub-partitioning algorithm #7814
• Add unit tests for GpuBatchSubPartitionIterator .
  (See Implement sub partitioning for large/skewed hash joins #7794 (comment))
  Fixed by PR Implement sub partitioning for large/skewed hash joins #7794

[firestarman]

close since all the sub tasks are done

[firestarman]

There is still one tracking issue #8057, but with low priority.