[python] Add ReadBuilder.with_filter (predicate pushdown)

[JunRuiLee]

Purpose

Second PR of exposing PyPaimon's DataFrame read path to Rust (follows #415). Refs #413.

Adds ReadBuilder.with_filter(predicate: dict) — converts a lightweight dict predicate into a Rust Predicate (resolving fields against the table schema) and pushes it into scan planning so plan() can prune splits.

table.new_read_builder().with_filter(
    {"method": "equal", "field": "id", "literals": [1]}
).new_scan().plan()

Supports the directly-translatable subset: equal/notEqual/lessThan/lessOrEqual/greaterThan/greaterOrEqual/isNull/isNotNull/in/notIn/and/or.

Out of scope (later PRs): read(splits) → Arrow data reading, pypaimon-side wiring.

Notes

• Schema is authoritative: any index/data_type in the dict is ignored; fields resolve by name against the current table schema.
• No partial pushdown: if any node (incl. a compound child) is unsupported, the whole with_filter fails — nothing is partially pushed.
• Unsupported operators (like/startsWith/endsWith/contains/not) and unsupported literal types (Date/Time/Timestamp/Decimal/Bytes, complex) → NotImplementedError. Unknown field / wrong literal count / None misuse / type mismatch / empty compound children → ValueError.
• Filter pushdown is conservative: successful conversion does not guarantee split-count reduction for every predicate; it enables read-side pruning where supported. Exact row filtering is left to a residual layer.

Tests

bindings/python/tests/test_read.py — filter conversion (eq/and/or/in/isNull/bool), partition pruning, and error paths (unsupported op/type, unknown field, wrong count, empty children).

Note on CI

CI will be red until #418 merges: current main does not compile (DataType::Vector not covered in arrow/format/row.rs, unrelated to this PR). This PR is built on main and will go green once #418 lands; happy to rebase then.

[JingsongLi]

+1