Optimize GPU radix sort with bitmask ranking and multi-element processing

[mvaligursky]

Optimizes the GPU compute radix sort with several algorithmic improvements, yielding ~70% faster sorting on a 10M-element benchmark (9.2ms → ~2.8ms).

Changes:

• Eliminate the local_prefix_sums GPU buffer entirely by computing local ranks in shared memory using per-digit 256-bit bitmasks and hardware countOneBits (popcount) — zero warp divergence
• Process 8 elements per thread (2048 elements per workgroup), reducing workgroup count 8× and shrinking the prefix sum hierarchy
• Simplify histogram pass to pure atomic counting — removes shared memory ranking that was previously interleaved
• Skip key buffer write on the last sorting pass (only values/indices matter for final output)
• Pre-fetch values before the workgroup barrier to overlap memory fetch with ranking computation
• Fix indirect dispatch mismatch by allocating separate dispatch slots for key generation (256 threads/workgroup) and sorting (2048 elements/workgroup)

API Changes:

• New export RADIX_SORT_ELEMENTS_PER_WORKGROUP from ComputeRadixSort module (used by compaction classes to compute correct dispatch sizes)

Performance:

• 10M-element 20-bit key sort: ~9.2ms → ~2.8ms (~70% improvement)
• Reduced GPU memory usage: one fewer storage buffer (local_prefix_sums) per sort instance