fix: preserve negative zero in JSONB floating-point values

[DarrenChangJR]

What this PR does and why

JSONB currently changes -0.0F and -0.0D into positive zero. The writer's
compact integer encodings use numeric equality, under which negative zero equals
zero, and discard the IEEE-754 sign bit.

This change recognizes negative zero using raw bits and bypasses only the
integer/zero compact encodings. It then uses the existing BC_FLOAT or
BC_DOUBLE representation, which preserves the sign. Positive zero, one, and
other integral floating-point values remain compact.

The behavior matters even though -0.0 == 0.0: raw-bit APIs distinguish the
values, division produces infinities with different signs, and Fastjson2's text
path already preserves negative zero.

Tests

NegativeZeroTest covers both float and double through:

• untyped scalar parsing;
• explicitly typed scalar parsing;
• primitive arrays;
• boxed arrays;
• primitive bean fields;
• boxed bean fields.

Validation performed:

• focused NegativeZeroTest;
• clean full mvn -pl core clean test suite: 7,977 tests, no failures;
• external JSONB invariant matrix: exactly 12 signed-zero violations removed,
  with all unrelated findings unchanged.

Checklist

• The change is limited to JSONB floating-point encoding.
• Existing compact encodings are preserved except where they lose the sign.
• Tests compare raw IEEE-754 bits, not numeric equality.
• No public API or documentation change is required.

中文说明

修改内容和原因

JSONB 当前会把 -0.0F 和 -0.0D 编码成正零。writer 的紧凑整数编码使用数值
相等判断，而负零与正零在该判断下相等，因此 IEEE-754 符号位会丢失。

本修改使用原始位识别负零，并且只让负零绕过整数/零的紧凑编码，继续使用现有
BC_FLOAT 或 BC_DOUBLE 表示。正零、1 和其他可紧凑表示的浮点整数保持不变。

回归测试覆盖 float/double 的无类型和显式类型标量、primitive/boxed 数组以及
primitive/boxed bean 字段，并用原始 IEEE-754 位比较结果。focused test、clean
full core 7,977 项测试和外部 JSONB invariant matrix 均通过。