Feature: Paged Attention Kernel

[ShlokVFX]

Implement a paged attention kernel to efficiently support long-context
inference by paging key/value caches.

The kernel should minimize memory fragmentation and reduce GPU memory
pressure while maintaining low latency.

Planned Benchmarks

• Latency vs sequence length
• Memory usage vs KV cache size
• Comparison against baseline attention

Learning Objectives

• KV cache paging strategies
• Memory locality and access patterns
• Tradeoffs between fragmentation and throughput