[python] Expose DataFrame-style scan/split planning API

[JunRuiLee]

Purpose

First PR of exposing PyPaimon's DataFrame read path to Rust. Refs #413.

Exposes the Rust core's ReadBuilder → new_scan() → plan() → DataSplit flow
through bindings/python:

• table.new_read_builder().with_projection([...]).with_limit(n).new_scan().plan()
• Plan.splits() returns Split objects that are picklable (opaque
  serde_json bytes), so they can cross process boundaries — groundwork for
  Ray distributed reads and a later read(splits) → Arrow step.

Out of scope (later PRs): with_filter / Predicate conversion, read(splits)
data reading, pypaimon-side wiring.

Notes

• with_projection is accepted and forwarded, but in PR1 it does not affect
  the planned splits — Rust core projection only applies to new_read() (data
  reading), not new_scan()/planning. It is staged here for the later
  read(splits) step. Tests assert only that planning succeeds.
• Split is an opaque payload: the serde encoding is an implementation detail,
  not a stable schema; only same/compatible-version round-trip is guaranteed.

Tests

bindings/python/tests/test_read.py — 7 tests: scan planning
(projection/limit/len) + pickle round-trip.