[python] Keep projected nested ROW predicates on native reads - #10637

Purpose

Keep existing PyPaimon nested ROW projection filters on Native scans and reads, rather than falling back when the projected leaf name is not a top-level table field.

Depends on apache/paimon-rust#1153. This PR remains draft until that core support is merged. Native CI continues building Rust main; no fallback or revision pin is introduced to bypass the dependency.

The existing Python API is unchanged: project profile.detail.score, then filter profile_detail_score using the builder's predicate builder.

Changes

• Retain each list projection's physical source path on predicates, including top-level columns renamed by collision suffixes. Preserve paths through index/literal rewrites and serialization.
• Pass physical roots and descendant paths to the Rust core predicate builder; Rust owns path validation, typed literals, exact evaluation and schema evolution.
• Align Python with the current Java NestedFieldTransform behavior: nested leaves cannot prune using an unrelated parent/scalar column's stats, partition/bucket keys or indexes.
• Fix a.b / real a_b collisions in Python exact filtering and Native transport. Preserve correct AND/OR pruning, row tracking and explicit partition-filter validation.

Tests

• Reproduced a real baseline Native fallback and guarded new Native tests against Python scan/read fallback.
• New coverage spans append/PK tables, batch/stream reads, NULL ancestors, comparisons/string predicates/ranges/membership, projection changes and LIMIT, field promotion/rename/drop-readd, empty/null-containing IN and NOT IN, reversed BETWEEN bounds, optimized LIKE, pickle, and collision cases with full statistics, separate files, partitioning, 8 buckets and AND/OR.
• Combined Python/Native regression: 998 passed, 2 skipped, 35 subtests passed, with 1021 Native plans and 1183 Native reads exercised.
• Additional review regressions: 96 collision cases, 26 partition tests and 354 existing predicate/projection/auth/index/bucket tests passed (groups overlap).
• Flake8 and git diff --check passed. Native tests ran against the built companion extension; extension rebuilds were completed before starting Native test runs.

Review follow-up

• Rebind filters by their physical paths after projection reordering, including collisions between a.b and a real a_b column. Reject a removed nested path rather than silently filtering the unrelated physical column.
• Keep top-level filter-only inputs in the Python adapter with aliases assigned together with output fields, preserving complete mixed AND/OR filters. Support ordinary physical predicates without a stored projection path. Native flattening reads only final output dependencies.
• Reject nested aliases in partition-only chunk-shuffle filters.
• Expand collision coverage to 192 cases and add 74 filter-only/row-ID cases covering append/PK, Native/Python, Arrow/batch-reader/iterator, duplicate projections, whole ROWs, NULL ancestors and overlapping PK commits.
• Final combined regression result: 1637 passed, 2 skipped, 39 subtests passed across 23 related Python suites, with 1355 Native plans and 1595 Native reads exercised. Extension compilation completed before Native tests started; no rebuild or install ran during the tests.