perf(table): prune parquet row groups by dictionary

[fallintoplace]

Summary

Use Parquet dictionaries during row-group pruning.

• EqualTo and IN predicates can skip dictionary-only row groups.
• Keeps the current stats and Bloom filter checks.
• Runs the dictionary check between stats and Bloom filters.
• Only trusts a dictionary when encoding stats show no plain fallback pages.
• Keeps the row group for missing metadata, unsupported types, malformed pages, or read errors.
• Groups dictionary predicates once per scan instead of once per row group.

Why

Min and max stats can be too broad for low-cardinality columns. A complete dictionary can prove that none of the requested values exist, so the row group does not need to be read.

Tests

• GOMAXPROCS=2 go test -p 1 -vet=off ./table/internal ./table -count=1
• Focused dictionary, Bloom filter, scan-planning, and malformed-page tests

Benchmark

Command:

GOMAXPROCS=2 go test -p 1 -vet=off ./table/internal -run '^$' -bench '^BenchmarkDictionaryRowGroupPruning$' -benchtime=1s -count=3

Workload: 16 row groups with 4,096 low-cardinality INT32 values each. The benchmark covers an absent target that prunes every group and a target present in one group.

• Without dictionary: 1.38 to 1.61 ms/op, about 1.62 MB/op.
• With dictionary, target absent: 0.65 to 0.81 ms/op, about 0.63 MB/op.