fix: canonicalize floating-point NaN values

[zjw1111]

Purpose

Linked issue: close #256

Canonicalize floating-point NaN payloads anywhere Java-compatible equality, hashing, or binary output requires a single representation.

This change centralizes float/double canonical bit conversion and applies it to:

• Literal hashing, Bloom filter hashing, and Hive bucket hashing.
• BTree key serialization.
• Variant float and double serialization.
• Global index result and indexed split score serialization.

Tests

• pre-commit run --files <all changed files>
• git diff --check
• Compiled all newly changed production and test translation units with the repository Debug compile commands and -Werror.
• MathTest.FloatingPointNaNCanonicalization
• FastHashTest.TestNaNCompatibleWithJava
• LiteralTest.*
• HiveBucketFunctionTest.TestFloatNaNCanonicalizationCompatibleWithJava
• HiveBucketFunctionTest.TestDoubleNaNCanonicalizationCompatibleWithJava
• GenericVariantTest.CanonicalizesFloatingPointNaN
• KeySerializerTest.CanonicalizesFloatingPointNaN
• GlobalIndexResultTest.TestSerializeCanonicalizesNaNScore
• IndexedSplitTest.TestSerializeCanonicalizesNaNScore

API and Format

No public API signature or protocol changes. Non-canonical NaN payloads are now written using the canonical Java-compatible NaN representation; existing canonical values and all non-NaN values are unchanged.

Documentation

No user-facing documentation changes.

Generative AI tooling

Generated-by: OpenAI Codex (GPT-5).

[lxy-9602]

+1