[Draft] Optimize radix sort for multi-work-group performance

[danhoeflinger]

This PR refactors the SYCL radix sort implementation to improve performance across multiple work-groups. The optimization focuses on kernel simplification, better memory access patterns, and improved work distribution.

Key Changes

Kernel Consolidation

• Removed the separate __radix_sort_reorder_peer_kernel
• Unified even/odd iteration logic to reduce kernel variants
• Eliminated compile-time kernel specialization overhead

Count Kernel Optimization

• Reorganized to support larger work-group sizes, previously limited to a single subgroup which resulted in poor occupancy
• Changed from _CountT (uint32) to uint8 accumulation in shared local memory (SLM) for bucket counters
• Improved SLM indexing with __index_views helper for better cache locality during counting phase
• Implemented 8-way unrolling with fully strided memory access for better coalescing
• Optimized reduction with tree-based aggregation where all work-items participate in parallel
• Reduced register pressure by having each work-item handle 4 radix states during reduction

Reorder Kernel Improvements

• Redesigned to support larger work-group sizes, previously limited to a single subgroup per work-group which resulted in bad utilization of the hardware

Code Simplification

• Removed complex peer prefix algorithm variants (subgroup_ballot, atomic_fetch_or, scan_then_broadcast)
• Removed workload-based tuning (_ONEDPL_RADIX_WORKLOAD_TUNING) in favor of simpler unified approach
• Net reduction of ~128 lines of code