[python] Support raw fallback for global indexes

[JingsongLi]

Summary

This PR ports the Java raw fallback behavior for global indexes to Python. It adds global-index.search-mode support and lets scalar global index scans and vector search include unindexed row ranges when configured with full or detail search modes.

Changes

• Add Python GlobalIndexCoverage and global-index.search-mode (fast, full, detail) with fast as the default.
• Extend Python GlobalIndexScanner with a separate unindexed_rows API and merge unindexed rows into scalar scan planning only when index scan succeeds.
• Add raw vector search splits and read-side raw vector scoring for Python vector search, including scalar pre-filter handling and partition pruning.
• Thread vector builder options into scan/read so raw vector scoring can use configured metrics.
• Add tests for scalar global index fallback modes, vector raw split planning, raw vector read merging, partition filtering, and metric selection.

Testing

• python -m pytest paimon-python/pypaimon/tests/global_index_test.py paimon-python/pypaimon/tests/vector_search_filter_test.py -q
• python -m pytest paimon-python/pypaimon/tests/range_test.py paimon-python/pypaimon/tests/partition_predicate_test.py paimon-python/pypaimon/tests/vindex_vector_index_test.py -q
• python -m compileall -q paimon-python/pypaimon/common/options/core_options.py paimon-python/pypaimon/globalindex paimon-python/pypaimon/read paimon-python/pypaimon/table/source
• git diff --cached --check

[leaves12138]

I reviewed this mainly against the Java raw-fallback implementation currently on master.

The Python changes look aligned with Java on the important semantics:

• global-index.search-mode keeps fast as the default, while full/detail compute unindexed row ranges from nextRowId / data-file row ranges.
• The normal data-evolution scan only ORs raw/unindexed rows when the scalar global index actually returns a result; unsupported index evaluation still falls back to the full table scan.
• Vector scan now builds indexed vector splits plus a raw split from vector-uncovered and scalar-filter-uncovered ranges, matching Java VectorScanImpl.
• Vector read uses empty indexed pre-filters when a scalar filter exists but no scalar index can evaluate it, so indexed vector search won't run unfiltered and pollute topK; raw search keeps the table filter applied when reading raw vectors.

Validation I ran locally:

• python -m pytest paimon-python/pypaimon/tests/global_index_test.py paimon-python/pypaimon/tests/vector_search_filter_test.py -q (53 passed)
• python -m pytest paimon-python/pypaimon/tests/range_test.py paimon-python/pypaimon/tests/partition_predicate_test.py paimon-python/pypaimon/tests/vindex_vector_index_test.py -q (24 passed)
• python -m pytest paimon-python/pypaimon/tests/global_index_test.py paimon-python/pypaimon/tests/vector_search_filter_test.py paimon-python/pypaimon/tests/global_index_evaluator_test.py paimon-python/pypaimon/tests/btree_thread_safety_test.py paimon-python/pypaimon/tests/global_index_update_action_test.py paimon-python/pypaimon/tests/index_manifest_write_test.py -q (89 passed)
• python -m compileall -q paimon-python/pypaimon/common/options/core_options.py paimon-python/pypaimon/globalindex paimon-python/pypaimon/read paimon-python/pypaimon/table/source

I also tried the full paimon-python/pypaimon/tests suite, but this local environment is missing optional deps such as parameterized and pypaimon_rust, so collection stops before executing the full suite.