Refactor CPU Partition dtype dispatch

[PhysicistJohn]

Summary

Replace CPU Partition's exhaustive dtype switch with
dispatch_all_types. The common dispatcher covers the same 14 current dtypes
and leaves the nth-element implementation, kth normalization and validation,
axes, output dtype, and diagnostics unchanged.

This removes 27 net lines from Partition::eval_cpu.

Testing

• Release CPU-only build with C++20 and global -Werror
• Baseline/candidate parity across all 14 dtypes, noncontiguous layouts, axes,
  kth boundaries, floating special values, and exact diagnostics (2,114
  cases); transcripts byte-identical
• Focused Python sorting tests (6/6 tests, 56/56 subtests)
• Native C++ suite (247/247 cases, 3,326/3,326 assertions)
• Global archive and linked-library export symbol sets unchanged
• pre-commit and git diff --check