[python] Add TableRead.read(splits) returning Arrow batches

[JunRuiLee]

Purpose

Third PR of exposing PyPaimon's DataFrame read path to Rust (follows #415 scan/plan/splits and #419 with_filter). Refs #413.

Adds the data-reading step both prior PRs deferred: ReadBuilder.new_read().read(splits) → List[pyarrow.RecordBatch].

b = table.new_read_builder().with_projection(["id"])
splits = b.new_scan().plan().splits()
batches = b.new_read().read(splits)   # -> List[pyarrow.RecordBatch]

Splits are picklable (from #415), so they can be planned in one process and read in another — groundwork for Ray distributed reads.

Notes

• Shared config: extracts apply_read_config (projection/limit/filter) so new_scan().plan() and new_read().read() apply the same pushdown consistently. Pure refactor, no behavior change to scan.
• Config snapshot: new_read() snapshots the builder's projection/limit/filter, so mutating the builder afterward does not affect an already-created read.
• Projection now applies: projection was a planning no-op in [python] Expose DataFrame-style scan/split planning API #415; it takes effect here on the read path.
• Eager collection: read() collects the full Arrow stream into a List[RecordBatch].
• Errors: a non-iterable argument or a non-Split element raises TypeError; empty splits return [] without entering the runtime. Off-GIL read via py.detach + runtime block_on, matching the existing plan() pattern.

Out of scope (later PRs): pypaimon-side wiring.

Tests

bindings/python/tests/test_read.py — 7 new tests: expected rows, empty splits, non-Split TypeError, pickled-split round-trip, projection applied, config-snapshot isolation, and a with_filter read smoke test. Full file 33/33 passing; cargo clippy clean.

[QuakeWang]

+1