Enhance the implementation of Random Sample Keys in kvrocks

### Search before asking

- [X] I had searched in the [issues](https://github.com/apache/kvrocks/issues) and found no similar issues.

### Motivation

Random sample keys is supported in kvrocks. When running commands like `spop`, scanning all data would be a bottleneck. Currently, the implementation is:
1. Get all sampled values
2. Random filtering K values
3. return

Unlike redis, it's complexity is always O(N), N is the element cardinal, which is extremly high. It's a bit hard to get "random" in kvrocks because we're based on rocksdb.

### Solution

There're some rules we can apply:
1. Maintaining a eq-depth histogram if random is frequently. This is the best one if random sample is frequently. However, we'll suffer from maintaining cost during write. TiKV using size based sampling to maintain the split range

So, instead, maybe we can enhance the implementation of current impl. For example:
1. Counting all "indices" of random values, and sorting them
2. Using iterator to get only these indices

The complexity is still O(N), but the performance might be enhanced.

### Are you willing to submit a PR?

- [x] I'm willing to submit a PR!

Start by locating kvrocks' current random-sample implementation used by commands such as spop and measure its all-data scan behavior. Compare the proposed index-counting and iterator approach with the existing behavior and Redis expectations; done requires an agreed implementation that improves sampling performance without changing command results.

Written by the indexing model from the issue text.