[improvement](parquet) Fuse dictionary selection planning

[Gabriel39]

What problem does this PR solve?

Predicate-only dictionary filtering currently expands definition levels and row selection into ColumnSelectVector, then scans that row-oriented state again to rebuild physical selection ranges and the selected NULL layout. Fragmented nullable filters spend significant CPU in these two planning passes before dictionary IDs are decoded.

Same-column conjuncts are combined into one dictionary-id bitmap before row decoding. Their incoming row selection therefore remains identity, which previously kept this common shape on the two-pass planner.

What is changed?

• Build physical dictionary selection ranges, selected NULL layout, and the filtered-row count directly while merging definition-level runs with the input filter.
• Pass the prepared plan into dictionary-index filtering instead of expanding and rescanning ColumnSelectVector.
• Plan identity-input dictionary selections directly when definition levels are materially fragmented.
• Use a bounded run-search fast path for no-NULL clustered filters, then switch to one linear scan after 32 runs.
• Keep small batches, dense nullable inputs, compact identity inputs, and sparse clustered nullable inputs on the existing planner.
• Add a self-validating 220-case dictionary-selection benchmark matrix, including 20 identity-input cases.

This PR does not change predicate decomposition, supported data types, decode strategy, or merge-read behavior. It is stacked on #66504.

Performance

Release build, 65,536 rows, CPU-pinned runs with 7 repetitions. The table reports median CPU time; lower is better.

Input selection	NULL layout	NULL rate	Existing planner	Selected planner	Change
Identity	Alternating	1%	11,066 ns	8,383 ns	-24.24%
Identity	Alternating	10%	98,943 ns	60,185 ns	-39.17%
Identity	Alternating	50%	580,955 ns	319,267 ns	-45.04%
Identity	Alternating	90%	87,733 ns	56,741 ns	-35.33%
Identity	No NULL	0%	1,451 ns	1,458 ns	+0.47%
Identity	Clustered	1%	1,488 ns	1,497 ns	+0.59%

The last two negative controls are routed to the existing planner; their sub-1% differences are measurement noise from identical code paths.

Validation

• 7 targeted ASAN Parquet dictionary and nullable-selection tests passed.
• All 220 dictionary-selection scenarios completed and self-validated identical physical ranges, selected NULL layout, and filtered-row counts.
• All affected C/C++ files pass clang-format 16.

[hello-stephen]

Thank you for your contribution to Apache Doris.
Don't know what should be done next? See How to process your PR.

Please clearly describe your PR:

1. What problem was fixed (it's best to include specific error reporting information). How it was fixed.
2. Which behaviors were modified. What was the previous behavior, what is it now, why was it modified, and what possible impacts might there be.
3. What features were added. Why was this function added?
4. Which code was refactored and why was this part of the code refactored?
5. Which functions were optimized and what is the difference before and after the optimization?

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: All Codex review accounts are usage-limited; earliest retry is 2026-08-08T03:32:00Z.
Workflow run: https://github.com/apache/doris/actions/runs/31020463260

Please trigger /review again after that time.

[Gabriel39]

run buildall

[Gabriel39]

/review

[github-actions]

Codex automated review failed and did not complete.

Error: All Codex review accounts are usage-limited; earliest retry is 2026-08-08T03:32:00Z.
Workflow run: https://github.com/apache/doris/actions/runs/31065828575

Please trigger /review again after that time.

[hello-stephen]

BE UT Coverage Report

Increment line coverage 65.23% (362/555) 🎉

Increment coverage report
Complete coverage report

Category	Coverage
Function Coverage	58.33% (24737/42411)
Line Coverage	42.59% (249504/585801)
Region Coverage	38.64% (197096/510130)
Branch Coverage	40.03% (90279/225521)