Use varint encoding for row count in encoding prefix (#657)

[tanjialiang]

Summary:

Change the Nimble encoding prefix row count field from a fixed 4-byte uint32_t to varint encoding. For typical small row counts (< 128 rows), this saves 3 bytes per encoding prefix.

The change is gated by a minor version bump (1 → 2) for backward compatibility:

• Write path: All encodings now use Encoding::Options{.useVarintRowCount = true} when index is enabled
• Read path: Files with version >= (0, 2) AND cluster index present decode with varint row counts; older files use the fixed 4-byte format

Key changes:

• Introduce NimbleFeatureVersion.h with NimbleVersion struct and NimbleFeatureVersionRange for centralized version-to-feature mapping. Future features follow the same pattern: add a version range, gate writer on current version, gate reader on file version.
• Bump kVersionMinor from 1 to 2 in Constants.h
• Thread useVarintRowCount=true through VeloxWriter, IndexWriter, and VectorizedStatistics encode paths
• Version-gate decode in VeloxReader, SelectiveNimbleReader, SelectiveNimbleIndexReader, ChunkedDecoder, IndexReader, and ReaderBase (stats deserialization)
• Fix tools (NimbleDumpLib, NimbleDslLib, EncodingLayoutTrainer, DumpNimbleTablet2DatabaseLib) to use varint-aware encoding traversal and row count reading
• Add useVarintRowCount parameter to traverseEncodings(), getStreamInputLabel(), and getEncodingLabel() in EncodingUtilities
• Update test utilities (TestUtils, VeloxWriterTest) to handle varint prefix parsing
• Update ~80 hardcoded physicalSize values in FieldWriterStatsTest to account for smaller prefixes
• Add NimbleFeatureVersionTest with version comparison, feature range, and constexpr tests

Differential Revision: D97174231

[meta-codesync]

@tanjialiang has exported this pull request. If you are a Meta employee, you can view the originating Diff in D97174231.