feat: integrate aisle for metadata-driven Parquet pruning

[belveryin]

Summary

Integrates aisle for metadata-driven Parquet pruning, enabling significant I/O reduction for selective queries by evaluating filter predicates against Parquet metadata before reading data.

Key Changes

• Row-group pruning: Skip entire row groups based on min/max statistics
• Page-level pruning: Fine-grained skipping using column/offset indexes
• Row filtering: Push predicates into Parquet reader for in-stream filtering
• Commit timestamp pruning: Skip SSTs entirely when min_commit_ts > read_ts

API Changes (Breaking)

// Before
Predicate::eq(ColumnRef::new("score"), ScalarValue::from(80_i64))
Predicate::gte(ColumnRef::new("age"), ScalarValue::from(18))

// After
Expr::eq("score", ScalarValue::from(80_i64))
Expr::gt_eq("age", ScalarValue::from(18))

• Predicate → Expr (re-exported from aisle)
• ColumnRef::new("col") → "col" (direct string)
• ScalarValue now from datafusion_common
• Method renames: gte → gt_eq, lte → lt_eq
• expr.not() → Expr::not(expr)

Dependencies

• Added aisle with row_filter feature
• Added datafusion-common for ScalarValue
• Upgraded arrow/parquet 56 → 57, fusio 0.5 → 0.6
• Removed tonbo-predicate crate

Implementation Details

1. SSTable writes now emit page-level statistics and offset indexes
2. Scan planning pre-computes row group/page selections and caches Parquet metadata
3. Residual evaluation rewritten with precise int/float comparisons (no precision loss)
4. Delete sidecars reuse metadata from plan phase

Test Plan

• Existing tests pass (226 tests)
• New tests for row-group/page pruning
• New tests for commit_ts pruning at plan time
• New tests for missing page index errors
• New tests for limit with residual predicates
• New tests for row filter with tombstones

Future Work

• Query metrics: Track pruning effectiveness (row groups/pages skipped, I/O bytes saved)
• Benchmarks: Measure I/O reduction and latency improvements for selective queries at varying selectivity levels (1%, 10%, 50%)

[ethe]

eprintln should be avoided, using logs or returning Result explicitly should be bnetter

[ethe]

The residual evaluation treats BloomFilterEq/InList as always true, which would produce incorrect results if these expressions were publicly available. Therefore, exposure should be restricted or true semantics should be implemented.

[ethe]

Basically the direction is good, there are some blockers we should fix before we merge this.

[belveryin]

addressed comments @ethe