partial_update: support sequence-group reads

[shyjsarah]

Purpose

Linked issue: close #583

Paimon Rust currently rejects existing merge-engine=partial-update tables that use fields.<sequence-fields>.sequence-group. This change adds read-side support so Rust and DataFusion can read Java/Flink-created sequence-group tables without falling back to incorrect basic latest-non-null semantics.

Brief change log

• Parse and validate single or composite sequence fields, protected fields, multiple independent groups, missing fields, primary-key conflicts, and repeated protected fields.
• Widen projected reads with the sequence fields required to merge projected or predicate-referenced protected fields, then remove those internal dependencies from the final projection.
• Merge each sequence group independently using Java-compatible ascending lexicographic comparison with nulls first.
  • Skip a group when its entire sequence tuple is null.
  • Let a newer or equal sequence tuple update the group.
  • Update sequence fields atomically and allow accepted protected fields to be set to null.
• Keep Rust CREATE/ALTER and writer paths fail-closed for sequence-group options.
• Continue rejecting sequence-group aggregation and remove-record/delete semantics, which are outside this PR.

Tests

• cargo test -p paimon --lib — 1797 passed, 1 ignored.
• cargo clippy -p paimon --lib --tests -- -D warnings.
• cargo test -p paimon-datafusion --test pk_tables — 52 passed.
• Added coverage for independent groups, composite tuples, all-null tuple skipping, partial-null ordering, null protected-field replacement, projection dependency widening, malformed options, and end-to-end projected reads.

API and Format

No public API or storage-format changes. This only expands read compatibility for existing partial-update tables. Rust table creation and writes remain unsupported for sequence groups.

Documentation

Updated the SQL documentation with the supported read-side behavior and explicit unsupported scope.

Review notes

The comparison behavior follows Java UserDefinedSeqComparator.create(..., true): ascending field order and nulls first. Equal sequence tuples use the later row in the existing merge order. The implementation intentionally does not add ownership restrictions beyond primary-key and repeated protected-field validation used by the Java path.

[JingsongLi]

+1