📝 docs: add GPU-Only XAttention guide with performance analysis

Add comprehensive documentation for GPU-only XAttention BSA mode:
- Architecture design and SparsePolicy interface
- Memory pre-allocation mechanism (alloc_policy_metadata)
- Performance analysis: 32K +15%, 64K +41% vs baseline
- CUDA Graph limitations explanation (variable seq_len in prefill)
- nsys profiling tools usage guide

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

CLAUDE.md

@@ -28,6 +28,7 @@ Nano-vLLM is a lightweight vLLM implementation (~1,200 lines) for fast offline L
 | [`docs/nsys_wrong_event_order_bug.md`](docs/nsys_wrong_event_order_bug.md) | 🐛 NSYS BUG: Ring buffer pipeline 触发 nsys 时间戳乱序问题的调试记录 |
 | [`docs/cpu_scheduling_latency_analysis.md`](docs/cpu_scheduling_latency_analysis.md) | ⚡ PERF: CPU 调度延迟分析，kernel 间隙来源，GPU 利用率优化方向 |
 | [`docs/bench_offload_results.md`](docs/bench_offload_results.md) | 📊 BENCH: CPU offload 性能测试结果，Full vs XAttention 对比 (32K/128K) |
+| [`docs/gpu_only_xattn_guide.md`](docs/gpu_only_xattn_guide.md) | 🚀 GPU-Only XAttention: 内存预分配、性能分析 (32K +15%, 64K +41%)、CUDA Graph 限制 |
 
 ## Rules Index