Unify global index construction behind a common table builder

[JingsongLi]

Purpose

Expose one global-index build API selected by index type, following Java's GlobalIndexer.create and GlobalIndexBuilderUtils. Sorting is an internal implementation choice; callers should not select separate sorted, full-text, Lumina or vindex table builders.

Brief change log

• Introduce Table::new_global_index_build_builder() and GlobalIndexBuildBuilder for all existing Rust global-index families: BTree, bitmap, multivalue, FM, full-text, Lumina and the five vindex types.
• Share column/type normalization, partition-filter validation and prepare/execute ownership. build() returns unpublished commit messages; execute() prepares and commits with the planned-snapshot guard.
• Make the specialized builders private implementations and migrate Rust callers, examples, DataFusion procedures and the Python binding/stubs. Preserve the Lumina-specific SQL procedure's type restriction.
• Use the table's REST-aware snapshot manager for vector preparation; honor per-build shard sizing and partition filters.
• Apply Java-compatible external global-index placement and DEIX source metadata to prepared generic indexes so the existing commit conflict checks also protect their inputs.
• Retain submitted or returned files on failed or uncertain commits; only failed private preparation cleans its own outputs.

Tests

• cargo test -p paimon --features fulltext --lib index_build --no-fail-fast: 142 passed, 2 previously ignored.
• cargo test -p paimon --lib global_index_build_builder: 54 passed, 1 previously ignored, including the build without the full-text feature.
• cargo test -p paimon-datafusion --features fulltext --test procedures index --no-fail-fast: 27 passed.
• cargo +1.98.0 clippy -p paimon -p pypaimon_rust --all-targets --features paimon/fulltext -- -D warnings: passed.
• cargo fmt --all -- --check and git diff --check: passed.
• Paired PyPaimon REST tests exercise real prepared full-text and all five vindex formats, both committers, partition conjunction, external placement, source conflicts and Python fast-search reads. Together with existing index-build tests: 153 passed and 19 subtests passed.

API and Format

Replaces the specialized public table factories/types with GlobalIndexBuildBuilder; the Python binding is a parameter/commit-message wrapper around core. The shared type-normalization helper is now named normalize_global_index_type.

Existing index formats and source metadata versions are unchanged. Generic builds now populate the already-supported DEIX source metadata.

Documentation

The new builder documents prepare vs execute, partition conjunction and supported column definitions; the Python stub and benchmark example use the common API.

[leaves12138]

Reviewed 6c30863f, together with apache/paimon#10487 at 7e986c78. No blocking findings.

The common entry point follows Java's type-selected factory model while keeping sorting and specialized writers internal. Ordered composite BTree definitions, type aliases, partition-filter conjunction and per-build shard options are preserved. build() returns unpublished messages; execute() retains the planned-snapshot guard and does not delete files after a failed or uncertain submission. The shared external placement and DEIX source metadata also preserve commit ownership and indexed-column conflict protection.

Independent local validation with Rust 1.94.1 and a freshly rebuilt Python extension:

• Core with full-text enabled: 4,010 passed, 6 ignored; formatting passed.
• Core without full-text: 54 relevant tests passed, 1 ignored, including the explicit unsupported-feature path.
• DataFusion index procedures: 27 passed.
• Python bindings: 417 passed.
• Paired REST/index-build suites: 153 passed, 19 subtests passed; 55 Native plans exercised.
• Related index/search regressions: 119 passed, 19 subtests passed.
• Additional reviewer tests: 34 passed for empty/all-NULL/sparse text shards, abort/rebuild ownership, effective option overrides and non-indexed-column updates; 3 passed against the previous Rust wheel to confirm early Classic fallback when the new factory is absent.

The real prepared full-text and all five vindex formats were read through existing Python fast-search paths and committed with both committers. Indexed-column changes reject the prepared messages without deleting their files; non-indexed-column changes remain committable.

Scope: Java alignment was checked against source and the DEIX byte contract, not a Java engine or production-storage interoperability run. The real Lumina-library test remains an existing ignore. Python intentionally retains its original nullable-vector construction path to preserve sparse relative row IDs.