[python] Read REST system tables with Native SystemRead

[JingsongLi]

Summary

Read the existing PyPaimon system tables through Rust core when Native reads are enabled on a supported REST FileStoreTable.

• Materialize metadata through Table.new_system_read(name).read(), returning a single PyArrow RecordBatch. The Native SystemRead also exposes read_type, with_projection, with_filter and with_limit.
• Store the full result in the existing Python SystemSplit. Existing in-memory projection/filter/limit, split reuse and pickle continue to work without recreating REST credentials or file IO on workers. No Native scan/plan/split classes are introduced.
• Delegate snapshots, schemas, options, manifests, files, buckets, tags, branches, file_key_ranges, table_indexes and consumers. Partitions retain their catalog-backed implementation.
• Match Java metadata semantics in Python: row-cast partitions, object-form keys/statistics, logical escaped file paths, actual file formats, physical-column/schema-evolved statistics, captured views and local metadata instants.
• Preserve declared schemas and empty results. Propagate Native read failures.

The changed metadata text follows Java. Logical escaped file paths can be passed through to_file_io_path before Python URI-based FileIO operations.

Dependency / CI

The companion apache/paimon-rust#1137 is merged. CI continues to build Apache paimon-rust main, with no fork or feature-branch pin.

Review follow-ups

• Format Python LTZ partition values from their epoch milliseconds and submillisecond fraction, without narrowing them into Arrow i64 nanoseconds. Preserve local-zone history/DST and submicrosecond fractions.
• Keep explicitly empty snapshot views empty when reading table_indexes; never pass the empty-view sentinel to scan(None), which requests latest.
• Construct metadata file paths from logical relative bucket paths, retaining file: and escaped components with data-file.path-directory.

Verification

• Native REST/system-table tests and existing system-table tests: 178 passed in UTC and 178 passed in America/New_York, using the companion extension.
• Added regressions failed before the fixes (9 failing cases) and passed afterward (14 focused cases). Shanghai timezone LTZ/path cases: 13 passed.
• Direct Native API tests cover filtering unprojected columns before projection/limit, read_type, zero limits and repeated/zero-column projections.
• Captured Python splits support another read configuration and pickle/offline reads without recreating REST.
• Prior global-index builder / SQL system-table regressions: 48 passed, 27 subtests passed. Escaped partition-source Native merge regressions: 24 passed.
• Flake8, license headers, Python 3.6 syntax and diff checks passed.
• Independent review of the simplified core/Python integration completed without P1/P2 findings.

[leaves12138]

LGTM