[python] Refine post-mask query filtering for authorized reads

[plusplusjiajia]

Purpose

Master (#10539) evaluates the full query predicate after column masking, but it applies every mask before the filter and resolves BLOB payloads before LIMIT. This PR keeps exact post-mask filtering and fixes the remaining performance and semantic gaps:

• Authorized reads evaluate every query predicate again before LIMIT and keep its columns until the final projection, because split reads may apply it loosely (nested projections, merged primary-key rows). Conjuncts that reference no masked column are still pushed down to scans and split reads.
• The filter tests a view in which only the masks it reads are applied; rows keep their raw values and all masks are applied after LIMIT. BLOB payloads are deferred until after filtering and LIMIT, so payloads that are output or read only by output masks are fetched for surviving rows only, while payloads read by filter masks stay eager.
• When a mask moves a column to another type family, such as a string mask on an INT column, comparisons, string predicates and null checks use the query-schema type, and values that do not fit it are treated as NULL, like SQL TRY_CAST. Numbers, times of one kind and nested columns are filtered as masked, so no precision is lost, and NOT IN with a NULL literal skips the conversion. NULL MAP values stay intact and _row_kind columns, including -D rows, pass through the filter.
• Streaming scans authorize filter columns that are not projected. Nested-field filter aliases authorize only their physical source column, and row-tracking fields are recognized.
• Scan pruning is rebuilt from the user predicate on every plan, so a later plan without masks restores full pruning. Only predicates on masked columns mark reads as filtered during reading, and chunk shuffling is rejected when authorization filters would resize chunks.

Tests

Adds regression coverage for append, primary-key and data-evolution tables, including partitions, projections, compound and OR predicates, LIMIT, streaming row kinds, chunk shuffling, NULL MAP values, type-changing masks, standalone equality and string predicates, selection-sensitive authorization for partition filters and hidden streaming filter columns, NOT IN with a NULL literal, masked values that do not fit the declared type (OR arms, NULL literals and null checks), numeric, timestamp and nested masks that change the column type, nested-field aliases, pushdown of unmasked conjuncts, OR predicates on unprojected columns under a nested projection, NULL comparisons after primary-key merges, and row-only authorization with a user filter. BLOB tests count payload reads: the post-mask LIMIT cases read 2 payloads instead of 10, including when an output-only mask reads a projected or unprojected BLOB.

• Python 3.11 query-auth, BLOB-deferral and predicate-parser suites: 464 passed.
• Python 3.11 read, filter, streaming, LIMIT, data-evolution and query-auth test files: 3051 passed; 15 failed only on missing optional modules (lance, vortex, duckdb), and 6 files need ray.
• Python-reader cases of the native query-auth, transform-filter and calendar-mask tests, run locally with the Rust-runtime skip removed: 198 passed. Native paths are covered by the Rust Plan CI job.
• Python 3.6 / PyArrow 6 query-auth filter suite: 194 passed; the MAP case is deselected because Polars 0.12.5 lacks nested-type support.
• Flake8 and git diff --check passed.

[JingsongLi]

Please rebase master.

[plusplusjiajia]

Rebased onto master. Since #10539 already filters after masking, this PR now keeps that exact post-mask filter and adds: only the masks the filter reads run before it, all masks run after LIMIT with BLOB payloads deferred, type-aware comparisons on masked values, and authorization of filter columns in streaming scans. Explicit partition filters on masked columns now use master's ValueError.