feat(file-index): support additional types and validate decimal scale

[zjw1111]

Purpose

Linked issue: #173

Extend file-index read/write support while keeping serialized data compatible with Java Paimon, and prevent silent DECIMAL index misfiltering at table entry points:

• Support FLOAT and DOUBLE in bitmap indexes.
• Support DECIMAL in BSI indexes.
• Support STRING, DECIMAL, and TIMESTAMP in range-bitmap indexes.
• Add a range-bitmap type adapter for logical-to-storage type and literal conversion.
• Add variable-length dictionary chunks for string keys and avoid eagerly reading chunk bodies.
• Cover canonical and payload NaNs, signed zero, infinities, and Java-generated bitmap bytes.
• Reject DECIMAL predicate literals whose scale differs from the schema in TableScan and TableRead creation.
• Allow different DECIMAL precision metadata when the value fits the field precision, and reject values that do not fit.
• Document that direct FileIndexReader and GlobalIndexReader Visit* calls must use the indexed field scale.
• Add regression coverage showing that mathematically equivalent DECIMAL literals with a different scale produce incorrect results in BSI, range-bitmap, and BTree readers.

Tests

• pre-commit run --files
• git diff --check upstream/main...HEAD
• Compiled the original file-index translation units with the repository Debug flags, including -Werror and ASan.
• BitmapIndexTest.* (14 tests)
• BitSliceIndexBitmapIndexReaderTest.:ChunkedDictionaryTest.:RangeBitmapFileIndexTest.:RangeBitmapTypeAdapterTest. (55 tests)
• SizeLimitAndCardinality/* (70 tests)
• PredicateValidatorTest.* (3 tests)
• BitSliceIndexBitmapIndexReaderTest.TestDecimalType and RangeBitmapFileIndexTest.TestWriteAndReadStringDecimalAndTimestamp (2 focused DECIMAL scale tests)
• Compression/BTreeGlobalIndexIntegrationTest.WriteAndReadDecimalCompactData/* (3 compression variants)

API and Format

No public API signatures are changed. TableScan and TableRead now reject DECIMAL predicate literals with a scale different from the schema, or a value that does not fit the field precision. Direct index-reader APIs retain their existing behavior and now document the same-scale requirement.

The bitmap, BSI, and range-bitmap formats are extended to the additional logical types without changing their format versions. Floating-point bitmap keys are verified against Java-generated V1/V2 bytes.

Documentation

Add public FileIndexReader and GlobalIndexReader notes requiring DECIMAL literals passed to Visit* to use the indexed field scale.

Generative AI tooling

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