[python] Expose DataFrame-style batch write + commit loop

[JunRuiLee]

Purpose

First PR of exposing PyPaimon's DataFrame write path to Rust. Refs #414.

Exposes the Rust core's batch write loop through bindings/python:

wb = table.new_write_builder()
write = wb.new_write()
write.write_arrow(batch)            # PyArrow RecordBatch, may be called repeatedly
messages = write.prepare_commit()
wb.new_commit().commit(messages)    # persists a snapshot; readable via SQL

CommitMessage (from prepare_commit()) is an opaque object passed back to commit() — same-process only for now; cross-process (pickle, for Ray worker→driver) is a later PR.

Out of scope (later PRs): serializable CommitMessage, overwrite, pypaimon-side wiring.

Notes

• write_arrow validates the incoming batch schema against the table schema and raises ValueError on mismatch (names + types; binary/fixed_size_binary tolerated), mirroring pypaimon's _validate_pyarrow_schema. No implicit casting — callers supply correctly-typed batches (e.g. INT → Arrow int32).
• A single commit_user is fixed per WriteBuilder and shared by new_write() and new_commit(), so the writer and committer agree (Paimon uses it for duplicate-commit detection).
• commit([]) (empty messages) is a no-op.

Tests

bindings/python/tests/test_write.py — 6 tests: write→commit→SQL read-back roundtrip, multi-batch, empty-commit no-op, schema-mismatch ValueError, non-message TypeError.

[JingsongLi]

This accepts LargeBinary/FixedSizeBinary as equivalent to Binary, but the batch is then passed through unchanged. The lower write path still uses the original Arrow arrays/schema: for example extract_datum_from_arrow only downcasts binary Paimon fields to arrow_array::BinaryArray, and append writes would also see the incoming schema rather than the table schema. So a batch that this check accepts can later fail with a type-mismatch (or write files whose Arrow schema differs from the table schema). Please either reject non-exact binary-family types here or normalize/cast the batch to the target Arrow schema before calling write_arrow_batch.

[JunRuiLee]

Good catch — confirmed the lower write path downcasts binary fields to arrow_array::BinaryArray only, so tolerating LargeBinary/FixedSizeBinary here was a false positive. Made type matching strict (reject rather than cast, consistent with write_arrow not casting) and added a regression test. Fixed in 2fe6b39.