Add CPU sort functions and optimize GPU radix sort with CUB

[keichi]

Summary

• Add CPU versions of sort functions (full_sort_cpu, partial_sort_cpu) using std::sort/std::partial_sort
• Replace custom GPU radix sort implementation with cub::DeviceSegmentedRadixSort for significantly better performance
• Add full_sort_with_scratch declaration to header and benchmark option
• Automatically select optimal sort implementation in ccm based on data size and available scratch memory

Performance

Benchmark results on GPU (time in ms, N×N matrix):

N	full_sort (Kokkos)	full_sort_with_scratch	full_sort_radix (CUB)
1,000	1.35	0.36	1.63
2,000	11.34	1.37	3.60
5,000	99.19	12.12	15.44
10,000	434.18	❌ memory limit	49.38
20,000	1,823.33	❌ memory limit	182.51

• For small sizes (N ≤ 5,000): full_sort_with_scratch is fastest
• For large sizes (N ≥ 10,000): full_sort_radix (CUB) is ~10x faster than default Kokkos sort

Test plan

• All 23 existing tests pass on CUDA build
• Benchmark verified performance improvements