Feature: RoPE (Rotary Positional Embedding) Kernels

[ShlokVFX]

Implement optimized RoPE kernels for attention, supporting both prefill
and decode paths.

Focus on minimizing compute overhead and improving data locality during
positional embedding application.

Planned Benchmarks

• RoPE overhead relative to attention
• Prefill vs decode performance
• Kernel fusion opportunities

Learning Objectives

• Positional encoding math
• Kernel fusion with attention
• Register and shared memory usage patterns