Fix binary NUMERIC-to-DOUBLE rounding

[ItsM0rty]

Fixes #604.

Unconstrained NUMERIC values and NUMERIC types outside DuckDB's DECIMAL limits are read as DOUBLE. The binary decoder currently accumulates their base-10000 digits with floating-point arithmetic, rounding at several intermediate steps. For example, PostgreSQL's 8629.433::numeric can be read as 8629.432999999999. Applying the display scale can also overflow while decoding finite values such as 1.5::numeric(1000,500).

This builds a decimal coefficient from the stored digits with exponent 4 * (weight + 1 - ndigits), then calls DuckDB's floating-point parser once. There is no need to expand implicit zero groups or apply the display scale with floating-point arithmetic. The handling of zero, NaN, infinity, and constrained DECIMAL values stays the same.

The new sqllogictest compares decoded values against PostgreSQL's float8 cast in COPY_BINARY, PLAIN_QUERY_BINARY, and PLAIN_QUERY_TEXT modes. It covers the reported values, halfway rounding, finite limits, subnormals, high display scales, NULL, special values, and NUMERIC arrays. Results are materialized before comparison so predicate pushdown cannot hide a decoder error.

Validation:

• The original extension fails the regression with 15 mismatches in the first 24-value dataset. The fix passes all 23 regression assertions.
• numerics_nan_inf.test, attach_types_decimal.test, and attach_types_numerics.test pass, with 295 assertions in total.
• Built DuckDB, the connector, CLI, and SQL test runner with Clang 18.1.3 on Ubuntu 24.04. Tests used PostgreSQL 16.15 and libpq 18.6. The local build used O1 without debug information to fit WSL memory; C++ API tests and unrelated extensions were disabled.
• clang-format 11.0.1 and whitespace checks pass.

CI also caught an existing giant-NUMERIC array expectation that recorded the old rounding error. The giant-NUMERIC and isolation-level expectations now use the corrected values (1e+32 and -1.2345678912345679e+26), checked against PostgreSQL's float8 and float8[] casts. Those two test files and the rounding regression pass locally with 50 assertions.

I haven't run the full connector suite, cross-platform tests, malformed-wire-input tests, or throughput benchmarks.

[staticlibs]

Thanks for the PR! Looks good to me.