[BENCH] Real-world query performance comparison

`examples/ctable/real_world.py` exists but is a fairly raw script. There is no self-contained example that clearly shows what performance gain SUMMARY indexes provide over a full scan, or how block size affects that gain — which is the most common question users will have after enabling auto-indexing.

Suggested work: Write `bench/ctable/summary_index_perf.py` that:
- Generates a synthetic CTable with a few million rows and numeric columns
- Runs the same where() query three ways: no index, SUMMARY at chunk granularity, SUMMARY at block granularity
- Prints a clean results table (rows scanned, time, speedup)
- Includes comments explaining the trade-offs

Ideally, it should work without any external dataset so it can be run immediately after install, but using an accessible dataset is also an option.

先阅读 examples/ctable/real_world.py，然后创建 bench/ctable/summary_index_perf.py。安装项目后运行自包含的 benchmark，并比较同一个 where() 查询在无索引、SUMMARY 的 chunk 粒度和 SUMMARY 的 block 粒度下的表现。完成的标准是：有一个清晰的表格报告扫描的行数、耗时和加速比，并附有解释权衡的注释。

由索引模型根据 Issue 内容生成。