perf(c/driver/postgresql): speed up decimal128 COPY encoding

[BOKJUNSOO]

Summary

This adds a fast path for PostgreSQL binary COPY encoding of Arrow
decimal128 values.

The existing path converts every Arrow decimal value to a decimal string, splits
that string into integer/fractional parts, then parses those parts back into
PostgreSQL's base-10000 numeric representation. For decimal128, we can avoid
that round trip: read the unscaled 128-bit integer directly, align it to
PostgreSQL's 4-decimal-digit groups, and write the numeric digits.

The fast path is guarded by __SIZEOF_INT128__. Platforms without native
unsigned __int128 support keep using the existing implementation.

decimal256 is unchanged.

Local benchmark

Environment:

• macOS arm64
• PostgreSQL 15 container
• Release build
• Python adbc_driver_postgresql.dbapi.adbc_ingest
• table schema: v numeric
• input: decimal128(38, 6), values Decimal(i) / Decimal(100)

Before, PyPI wheel adbc-driver-postgresql==1.11.0:

100k rows: ~2.166s, ~46k rows/s

Before, local unpatched Release build from the ADBC source tree:

100k rows: ~0.831s, ~120k rows/s

After this patch:

100k rows: ~0.058s, ~1.73M rows/s
200k rows: ~0.097s, ~2.05M rows/s
1M rows:   ~0.486s, ~2.06M rows/s

Testing

PKG_CONFIG_PATH=/opt/homebrew/lib/postgresql@14/pkgconfig \
  cmake -S c -B build-local \
  -DADBC_DRIVER_POSTGRESQL=ON \
  -DADBC_BUILD_SHARED=ON \
  -DADBC_BUILD_STATIC=OFF \
  -DADBC_BUILD_TESTS=ON \
  -DADBC_BUILD_BENCHMARKS=OFF \
  -DADBC_WITH_VENDORED_FMT=ON \
  -DADBC_WITH_VENDORED_NANOARROW=ON \
  -DCMAKE_BUILD_TYPE=Release

cmake --build build-local --target adbc-driver-postgresql-copy-test -j 8

ADBC_POSTGRESQL_TEST_URI='postgresql://postgres:password@127.0.0.1:9900/postgres?sslmode=disable' \
  build-local/driver/postgresql/adbc-driver-postgresql-copy-test

Result:

[==========] 73 tests from 5 test suites ran.
[  PASSED  ] 73 tests.

I also checked PostgreSQL roundtrips for scale 2/6/8/18, negative values,
small fractional values, and 38-digit decimal128 values.

[lidavidm]

CC @Mandukhai-Alimaa for review

[Mandukhai-Alimaa]

LGTM.

[lidavidm]

It appears we need to use __extension__ __int128 or similar to make GCC happy: https://stackoverflow.com/a/54649367

[BOKJUNSOO]

Thanks. Fixed in d929cfc by defining __extension__ using UInt128 = unsigned __int128; under the existing __SIZEOF_INT128__ guard and using the alias throughout the fast path. I verified the actual writer header with GCC 14 using -std=gnu++17 -Wpedantic -Werror; the PostgreSQL COPY test suite also passes (73 tests).

[lidavidm]

@Mandukhai-Alimaa want to take a look as well?

[lidavidm]

Oh wait, you already did.