Avoid allocating a digits array when writing decimal as numeric

Fixes #6728

DecimalNumericConverter<T>.Write stackalloc'd MaxDecimalNumericDigits (8) shorts. For a decimal's 3 uints, PgNumeric.Builder.Create reserves (3 * 10 + 3) / 4 + 1 = 9 digits, and falls back to new short[9] when the destination is smaller. So every non-null value written through this converter allocated 48 B, and the stack buffer was never used. That covers decimal, plus the byte/short/int/long/float/double to numeric mappings.

Changes:

• PgNumeric.Builder: the bound expression moves into GetDestinationLength(bitsLength), which Create now uses. A new internal const int DecimalDestinationLength (9) uses the same expression, and AssertInvariants checks in debug builds that the two agree.
• DecimalNumericConverter<T>.Write stackallocs DecimalDestinationLength.

MaxDecimalNumericDigits stays 8. That is the real maximum for a decimal, and ToDecimal and DecimalBasedMaxByteCount depend on it. A simpler MaxDecimalNumericDigits + 1 at the call site, or a tighter bound for decimals inside the builder, would also work. I preferred not to change the shared conversion code.

Tests (NumericTests):

• New ReadWriteCases with 29 significant digits at scales 28, 27, 26, 5, 4, 3 and 1, positive and negative, so the existing read/write round-trip tests cover values that use all 8 numeric digits, with and without the scale-remainder digit.
• Builder_from_decimal_uses_destination ([IssueLink] to Writing a decimal as numeric allocates a short[9] per value #6728) builds PgNumeric.Builder from a DecimalDestinationLength buffer. It checks that the digits land in that buffer (no array allocated) and that the value round-trips exactly. With the buffer set to 8, 68 of the 69 cases fail; only the value 0 passes, as it writes no digits. It tests the builder, not the converter's call site; the benchmark below covers that.

TypeHandlers.Numeric.Write (existing benchmark, MemoryDiagnoser), ShortRun, .NET 10.0.14, 4-core Xeon 2.1 GHz VM:

All 15 benchmark values go from 48 B to 0 B. Mean times were 58-109 ns before and 49-99 ns after, but the ShortRun error bars overlap, so this PR claims only the allocation change.

Types.NumericTests and Types.NumericTypeTests pass on net10.0 in Debug and Release (372/372) against PostgreSQL 17.

For 10.0.x: the same change applies to hotfix/10.0.4, where the method is still WriteCore. git apply -3 merges the source files but leaves a conflict in NumericTests.cs.

Unrelated to #6509 (read-side ToDecimal bug).