Improve performance for hash partition operator

[tgujar]

Description

This change adds a staged scatter path for fixed-width hash partitioning. It computes compact row metadata once, then copies fixed-width output columns directly in partition order instead of materializing the whole result through a gather map. Variable-width columns and validity masks continue to use the existing gather machinery.

The implementation includes these optimizations:

• pack each row's partition ID and local partition offset into one 32-bit metadata value when the
  bit widths allow it, with the existing representation retained as a fallback;
• hash compatible fixed-width keys through a specialized row hasher and remove redundant global
  partition-total storage;
• stage partition offsets in shared memory and copy 1-, 2-, 4-, 8-, and 16-byte values directly,
  including vectorized 16-byte copies;
• select the staged-scatter CTA size and rows per thread from the available shared memory;
• fuse fixed width column copies to avoid repeated metadata loads from GMEM for each column
• use streaming loads for one-pass input data and group output columns into L2-sized batches to
  reduce scattered output sector writes;
• build one gather map only when variable-width columns or validity masks need it, then reuse it for
  all remaining materialization work.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[tgujar]

Some perf charts on RTX 6k PRO Blackwell