Integrate FileIndex pruning into raw data-file reads

[QuakeWang]

Search before asking

• I searched in the issues and found nothing similar.

Description

Issue #34 was closed by PR #764 after landing the Java-compatible FileIndex outer format, Bitmap and Bloom Filter readers, built-in dispatch, and predicate/result composition.

However, these components are currently used only by tests. The production DataFileReader path does not consume DataFileMeta.embedded_index or .index sidecars, so normal Rust reads cannot benefit from FileIndex pruning on Java-written tables.

This task should integrate the existing FileIndex components into raw data-file reads, following Java Paimon's FileIndexEvaluator and RawFileSplitRead.

Proposed scope

• Support file-index.read.enabled, defaulting to true as in Java.
• Resolve FileIndex data with Java-compatible precedence:
  • Prefer embedded_index.
  • Otherwise use one .index entry from extra_files, resolved with aligned_file_path.
  • Return Remain when no index exists.
  • Fail on multiple sidecars only when no embedded index is present.
• Add a bytes-backed reader for embedded FileIndex data.
• Evaluate predicates against the data file's schema:
  • Rebind fields by field ID to the file-level name, index, and type.
  • Safely devolve literals for schema evolution.
  • Fall back to Remain when predicate devolution is unsafe.
• Skip unsupported index identifiers before reading their payloads, while continuing to use supported indexes from the same container.
• Apply FileIndexResult before opening the data file:
  • Remain keeps the existing path.
  • Skip avoids reading the data file.
  • Selection is intersected with split row ranges and deletion vectors.
• Keep the original data predicates as residual filters. FileIndex is pruning only and must not replace exact row-level filtering.
• Use FileIndex selections only when the file row count fits the non-negative i32 range used by the Java-compatible Bitmap format.
• Preserve the existing conservative error behavior: structural format errors are propagated, while unsupported operations and lazy evaluation failures fall back to Remain.

Tests

Cover:

• Embedded and sidecar indexes.
• Regular and external-path sidecar resolution.
• Bitmap selection and Bloom Filter false positives.
• Intersection with deletion vectors and split row ranges.
• Schema rename, reorder, added columns, and type promotion.
• Unsupported identifiers and the i32 row-count boundary.
• Identical query results with FileIndex enabled and disabled.

Out of scope

• FileIndex writer and creation-option integration.
• Additional index implementations, TopN, and limit-only evaluation.
• Nested map-expression extraction.
• DataEvolutionReader and merge-reader integration.
• New public FileIndex APIs.

This is a read-side follow-up to #34 and #764. Writer integration and additional reader or index types should be tracked separately.

Willingness to contribute

• I'm willing to submit a PR!