[python] Use Native scan statistics in explain

[JingsongLi]

Purpose

Use Native planning counters in PyPaimon's existing explain flow and exercise Java-compatible fixed-bucket pruning through Native scans.

Depends on apache/paimon-rust#1140. This PR remains draft until the API is available on Rust main, which Native CI builds.

Changes

• Exercise Native same-key OR, repeated IN intersections, correlated partition predicates and composite bucket keys through the existing scan/read API.
• Align Python's selector with Java's conservative treatment of pending buckets and predicates which become AlwaysFalse after partition substitution. Partition filtering rejects nonmatching partitions.
• Native scans always collect lightweight counters. Use scan.plan() and plan.stats(); remove the collect_stats bridge flag and the Native with_stats call. Preserve early manifest/sidecar pruning and LIMIT stopping.
• Keep snapshot metadata on Plan. Preserve counters through pinned snapshots, incremental windows, streaming conversion, post-processing and captured authorization without rerunning the Python scanner.
• Keep a small Native stats handle and defer conversion to Python counters until Plan.stats() is called. Cache the converted result. Authorization wrapping preserves lazy conversion, and pickle materializes Python scalars instead of serializing the process-local stats handle.
• Reconcile final split/file/byte totals and deferred LIMIT metadata with the returned Python Plan when statistics are first requested, retaining Native decode/prune counters. Empty incremental windows use the same complete zero-counter shape as an empty core Plan.
• Preserve scan_with_stats() for both planners. Render actual Native work through ExplainResult.planning_counters. Full before/after funnels and Python identity sets are unavailable for Native because skipped metadata is not decoded.

Rust owns core counter collection, schema statistics, Data Evolution projection and split planning. Python converts and displays the result, reconciling final totals after its existing post-processing and deferred LIMIT. No new format support.

Verification

• After the review fixes, expanded Native regression: 516 passed, 13 skipped, 122 subtests passed. Its remaining authorization mock was updated to use the complete Native counter factory; both affected unit suites then passed all 119 cases (46 subtests), including that authorization case.
• Added 12 REST regression cases covering shard/slice with deferred LIMIT and empty timestamp windows: all 12 failed before the fix and passed afterward. Added a unit case preserving work counters, saturating final bytes, excluding unknown sizes and retaining zero LIMIT. Statistics remain lazy, cached and serializable.
• Native regressions verify counters on ordinary plans, lazy conversion and reuse, pickle/JSON serialization, LIMIT stopping, historical/empty snapshots, cross-schema netting, Data Evolution projection, captured grants, and no data-file reads during explain.
• Changed Python files passed repository Flake8 configuration and Python 3.6 syntax parsing.
• Companion Rust verification: 2105 table-layer tests, 168 binding read/streaming tests and 6 DataFusion statistics tests passed; Rust formatting and CI's Rust 1.98.0 workspace Clippy passed.