perf(cpp): push down multi-value aligned offsets

[ColinLeeo]

Motivation

Multi-value aligned reads accepted row offsets at the upper query layer, but the offset was not pushed down into the chunk/page read path. As a result, rows before the requested offset could still force unnecessary page decompression and decoding.

Changes

• Add chunk-level offset skipping for multi-value aligned series scans when the time chunk and all value chunks have matching row counts.
• Allow multi-value aligned page reads to accept row_offset.
• Add page-level offset skipping in the serial multi-value aligned page loop using the time page statistic count.
• Preserve partial-page residual offsets so the row reader can consume the remaining in-page offset correctly.
• Push dense aligned/no-filter table row offsets into the shared multi-value SSI.
• Synchronize residual SSI offset back to the row reader to avoid double skipping.
• Add targeted tests for partial-page residual offsets, whole-page offset skipping, and whole-chunk offset skipping.

Performance

Tested on a 6-file time-series dataset with 264 logical series and about 996 million points. The benchmark repeatedly reads fixed 2880-point windows with randomized offsets.

Before this change:

• ~34 ms/read
• ~29 reads/s

After this change:

• ~0.65 ms/read
• ~1538 reads/s

Overall speedup is about 52x for this offset-heavy random window read workload.

For high-offset reads on the longest series, the 90% offset probe improved from about 410 ms to about 0.99 ms.

Validation

cmake --build cpp/build/test-noantlr --target TsFile_Test -j 8

cpp/build/test-noantlr/test/lib/TsFile_Test \
  --gtest_filter='TsFileReaderTest.*'

cpp/build/test-noantlr/test/lib/TsFile_Test \
  --gtest_filter='TableQueryByRowTest.*Offset*:TableQueryByRowTest.DenseSingleDeviceSsiLevelPushdown:TableQueryByRowTest.DenseAlignedNullsMustUseTimeRowCount'

Results:

- TsFileReaderTest.*: 14 passed, 1 disabled
- table row offset related tests: 15 passed