feat: add BloomFilter, CRC32C, MurmurHash, varint utilities, and DeltaVarintCompressor

[lxy-9602]

Purpose

No Linked issue.

Introduce hashing, checksum, and compression utilities:

• BloomFilter / BloomFilter64 — probabilistic membership filters with 32-bit and 64-bit hashing (bloom_filter.h/cpp, bloom_filter64.h/cpp)
• CRC32C — CRC32-Castagnoli checksum with SSE4.2 hardware acceleration fallback (crc32c.h/cpp)
• MurmurHashUtils — MurmurHash3 implementation for byte arrays and MemorySegments (murmurhash_utils.h)
• VarLengthIntUtils — variable-length integer encoding/decoding (var_length_int_utils.h)
• DeltaVarintCompressor — delta + varint compression for integer sequences (delta_varint_compressor.h/cpp)

Also adds xxhash_test.cpp and data_define_test.cpp test coverage and complete testharness.h/.cpp.

Tests

• bloom_filter_test.cpp, bloom_filter64_test.cpp — filter add/contains, false-positive rate
• crc32c_test.cpp — checksum correctness across input sizes
• murmurhash_utils_test.cpp — hash consistency and known-vector checks
• var_length_int_utils_test.cpp — varint round-trip encoding
• delta_varint_compressor_test.cpp — compression/decompression round-trip
• xxhash_test.cpp, data_define_test.cpp

API and Format

Documentation

Generative AI tooling

Migrate-by: Aone Copilot (Claude)

[lxy-9602]

Thanks @ChaomingZhangCN for contributing the BloomFilter and CRC32C implementations — migrated as part of this batch. 🎉

[leaves12138]

Thanks for the migration. I found one blocker: DeltaVarintCompressor currently relies on Java-style signed wraparound in the extreme-value case, but the C++ implementation performs signed int64_t subtraction/addition, which is undefined behavior and can fail under UBSan or compiler optimization. Please implement the delta/reconstruction in a well-defined modulo-2^64 domain so it remains Java-compatible without UB.

[leaves12138]

This subtraction can overflow for the tested {INT64_MIN, INT64_MAX} case. Java long arithmetic wraps, but C++ signed overflow is undefined behavior. Please compute the delta with well-defined modulo semantics, for example via unsigned intermediates, before ZigZag encoding.

[leaves12138]

Same issue during reconstruction: for the extreme-value case, INT64_MIN + (-1) relies on signed overflow to become INT64_MAX. Please reconstruct using well-defined modulo semantics instead of signed overflow.

[leaves12138]

Reviewed the latest head. The previous DeltaVarintCompressor signed-overflow blocker has been fixed with unsigned delta/reconstruction, and I did not find new blockers in this pass.