refactor(query): combine short-query and page-aware execution

[dbsid]

I hereby agree to the terms of the CLA available at: https://docs.databend.com/dev/policies/cla/

Summary

Draft integration PR for the reviewed changes in #20165 and #20166.

This branch combines:

• short-query execution overhead reductions across session, logging, privilege, MySQL runtime, and pipeline executor paths;
• page-aware lazy RowFetch for Fuse Parquet reads, including bounded page-range work, cache isolation, and full-block fallbacks;
• the focused correctness and CI race fixes already present on both source PRs.

The purpose of this draft is to continue profiling the combined path and reduce the 5-concurrent p99 latency of Q1, Q2, and Flex below 10 ms. The target is not claimed yet: the prior #20166 c5 lazy benchmark reached p99 values of 34.66, 38.68, and 39.57 ms for Q1, Q2, and Flex respectively.

Composition

• Base: origin/main at 7dc41477f7.
• refactor(query): reduce short-query execution overhead #20165: b934267071ea7c77cac027e63f7d56950844108b.
• feat(query): make lazy row fetch page-aware #20166: 643c2fc22b2452be6e2400068e1c8f93766f9d3d.
• Combined branch uses explicit no-fast-forward merge commits so both source histories remain reviewable.

Benchmark Contract

Driver: https://github.com/dbsid/hbx-web3

• EC2 i4i.4xlarge, single query node, S3 in us-west-2, native release build;
• Q1, Q2, and Flex with 5 concurrent clients and lazy_read_threshold=1000;
• 3,000 measured requests per query after one warm round, second round authoritative;
• deterministic seed 2026071305, max_threads=2, max_storage_io_requests=16;
• distributed pruning and query-result cache disabled; planner cache held constant;
• zero errors and ordered-result comparison against lazy_read_threshold=0 are required for every claimed result.

Profiling Plan

1. Establish a red-capable c5 p99 loop on the combined binary.
2. Measure protocol, executor scheduling, session/logging, RowFetch, Parquet page reads, cache, and result serialization boundaries.
3. Rank falsifiable hypotheses and change one variable at a time.
4. Retain only optimizations with focused regression coverage and a measured causal improvement.
5. Repeat the exact benchmark until Q1/Q2/Flex authoritative p99 are all below 10 ms, or document the measured blocker without claiming completion.

Current Validation

• Both source PR validation suites passed before composition.
• The combined merge has no conflicts and git diff --check passes.
• New profiling and optimization results will be added here as they are produced.

Tests

• Unit Test (inherited from refactor(query): reduce short-query execution overhead #20165/feat(query): make lazy row fetch page-aware #20166)
• Logic Test (inherited from refactor(query): reduce short-query execution overhead #20165/feat(query): make lazy row fetch page-aware #20166)
• Benchmark Test (combined c5 p99 target pending)
• No Test - Explain why

Type of change

• Bug Fix
• New Feature
• Breaking Change
• Documentation Update
• Refactoring
• Performance Improvement
• Other (please describe):

---

This change is