Cache simulator

We want to study the how different evict/promote policies impact performance. In order to do that, we need a cache trace and a cache simulator that can replay the trace.

To collect cache trace, we need to instructment all the cache get operations:

1. Get arrow array with filter https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/cache/mod.rs#L266
2. Evaluate selection with predicate: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/cache/mod.rs#L206

We will get a trace like this:

1. Get(RowGroup id, column id, row id, size)
2. Get(RowGroup id, column id, row id, size)
3. Get(RowGroup id, column id, row id, size)

One we have the trace, we want to build a cache simulator which explores:

1. Eviction policy: when and which batch to evict?
2. Promotion policy: should we insert the missed batch to cache?
3. How does cache size impact performance

Simulator will have one interface: get(RowGroup id, column id, row id) -> Option.
On cache hit, it returns Some(V); on miss it returns None.

Internally, the simulator decides whether to insert the (RowGroup id, column id, row id) to cache.

Reference: https://blog.xiangpeng.systems/posts/parquet-pushdown/

Start with the cache get operations in src/liquid_parquet/src/cache/mod.rs at the linked locations around lines 206 and 266, and read the referenced cache-trace design. Define how traces record row-group, column, row, and size values, then determine the simulator behavior for get(RowGroup id, column id, row id) across eviction, promotion, and cache-size policies. Done means a trace can be replayed and policy performance can be compared.