Feature: Decode Fusion Kernel

[ShlokVFX]

Implement a fused decode kernel combining logits processing, sampling,
and token selection into a single GPU launch.

This feature aims to reduce kernel launch overhead and improve
token-level throughput during autoregressive decoding.

Planned Benchmarks

• Tokens/sec improvement
• Kernel launch count reduction
• Latency per decode step

Learning Objectives

• Kernel fusion techniques
• Launch overhead analysis
• Decode path optimization