Feature: Custom Sampling Kernels

[ShlokVFX]

Implement custom token sampling kernels for inference, including
top-k, top-p, and temperature-based sampling.

The goal is to reduce CPU-GPU synchronization and avoid framework-level
sampling overhead.

Planned Benchmarks

• Sampling latency per token
• End-to-end decode latency
• Comparison with framework sampling

Learning Objectives

• Parallel prefix sums
• Warp-level reductions
• Sampling stability and numerical behavior