refactor: split statistics sketch storage and add RocksDB diagnostics

[KKould]

What problem does this PR solve?

• The previous statistics layout stored the Count-Min Sketch too close to hot table/index data, which made RocksDB locality worse and showed up as a large TPCC regression.
• Statistics cache misses were also repeatedly falling back to storage, including negative lookup cases, which made the regression harder to reason about.
• We also lacked phase-level RocksDB diagnostics, so it was difficult to tell whether cache misses came from load, analyze, or the measured TPCC workload.

Issue link:

What is changed and how it works?

• Split Count-Min Sketch persistence out of the statistics root and store it as sketch meta + sketch pages.
• Keep histogram metadata and buckets as the eager statistics payload, and load sketch data lazily only when collect_count actually needs it.
• Change StatisticsMetaCache to cache both positive entries and negative lookup results via Option, so repeated misses do not keep hitting storage.
• Update ANALYZE, storage codecs, and transaction APIs to read/write the new statistics layout.
• Add richer RocksDB metrics, including data/index/filter cache hit/miss counters and TPCC phase snapshots, so locality regressions are easier to diagnose.
• Bump crate version to 0.1.8 for the next publish.

Code changes

• Has Rust code change
• Has CI related scripts change

Check List

Tests

• Unit test
• Integration test
• Manual test (add detailed scripts or steps below)
• No code

Manual test steps:

• cargo test test_analyze --lib
• cargo check -p tpcc -p kite_sql
• cargo run -p tpcc --release -- --backend kite --measure-time 60 --num-ware 1 --rocksdb-stats --path /tmp/kitesql_tpcc_phase_metrics
  • Observed TPCC stayed in the restored ~24k TpmC range and phase metrics showed misses were dominated by data blocks during the measured phase.

Side effects

• Performance regression: Consumes more CPU
• Performance regression: Consumes more Memory
• Breaking backward compatibility

Note for reviewer

• The core of this PR is the sketch-storage split and the cache/load path changes around statistics.
• The RocksDB TPCC diagnostics are included to validate the locality impact of the refactor and to make future regressions easier to spot.