[mosaic] Expose column-batch reads and writer extension hooks

[jianguotian]

Purpose

Mosaic reads Arrow batches internally, but currently exposes row iterators to callers. Expose those batches through Paimon's existing vectorized reader contract so downstream engines can process columns without unpacking rows.

• ArrowBatchReader and MosaicRecordsReader expose independent batch wrappers and release Arrow buffers per consumed batch, preserving projection, prefetch and row-tracking fallback.
• MosaicRecordsWriter exposes its conversion buffer, flush and failed-state check to subclasses. Failed native writes reject further input and are not retried during close.
• MosaicWriterFactory adds an overridable creation hook while retaining option validation.

Existing bundle interfaces, column-to-Arrow conversion and native Mosaic writing are reused. MosaicRecordsWriter continues to use the row fallback for ordinary VectorizedBundleRecords; consumers can implement batch-write adapters through the new hooks.

Tests

Passed clean Maven verification on JDK 8 (linux/aarch64): 131 tests, zero failures, errors or skips. Checkstyle, Spotless and Enforcer checks were enabled.

mvn -B -pl paimon-arrow,paimon-mosaic -am \
  -Dtest=ArrowBatchReaderTest,Mosaic*Test -DwildcardSuites=none \
  -DfailIfNoTests=false -Dsurefire.failIfNoSpecifiedTests=false clean verify

Coverage includes prefetch, projection, row tracking across row groups, deletion vectors, batch lifetime, native-write failures, external writer subclass round trips and factory option validation.

[JingsongLi]

+1