Feature/query by row

[hongzhi-gao]

Summary

Add queryByRow(paths/table, offset, limit) for both tree and table model. Results are equivalent to “full query then skip first offset rows and take at most limit rows,” but offset/limit are pushed down so that Chunk/Page-level skipping avoids decoding unnecessary data where possible.

Changes

Tree model

• API: TsFileReader::queryByRow(path_list, offset, limit) / TsFileTreeReader::queryByRow(devices, measurements, offset, limit).
• Pushdown: Single-path: set_row_range(offset, limit) on SSI → Chunk/Page skipped by count. Multi-path: offset/limit applied in merge loop; min_time_hint used to skip stale Chunks.
• Tests: Correctness (no offset/limit, offset only, limit only, offset+limit, boundaries, multi-path merge) + QueryByRowFasterThanManualNext (timing: queryByRow faster than full query + manual next, 5% tolerance).

Table model

• API: TsFileReader::queryByRow(table_name, column_names, offset, limit).
• Pushdown:
  • Device: skip whole device when remaining_offset >= device_row_count (Dense (in this codebase) means: within one device, every queried column has the same number of rows and the same timestamps.).
  • SSI: when dense, set_row_range(offset, limit) on each column’s SSI → Chunk/Page skip by count.
  • TsBlock: when sparse or not fully consumed at SSI, offset/limit applied in merge loop.
• Tests: Correctness (single/multi device, offset/limit, boundaries, equivalence with manual skip, SSI-level pushdown) + QueryByRowFasterThanManualNext (same timing check as tree).

Review focus

• Semantics: queryByRow(offset, limit) matches “full query + skip offset + take limit” (existing equivalence tests).
• Performance: New timing tests require queryByRow to be no slower than manual next within 5% (min of 5 runs); confirms pushdown is used in practice.