ranges: trim interval join build payload

[mwiewior]

Partially addresses #33

Summary

• build interval indexes directly in algorithm-specific buffers instead of first materializing BioInterval
• replace the callback-based probe API with collect_matches(..., &mut Vec<u32>)
• fill probe-side row indices directly while probing instead of RLE -> expand_probe_indices
• project the concatenated build-side batch down to only the left columns needed by the join output or residual filter
• add examples/interval_join_bench.rs to benchmark hash join vs interval join algorithms on count-heavy and wide-output overlap queries

Why

IntervalJoinExec still spent a disproportionate amount of time in build-side materialization on count-style overlap queries. The first PR branch concatenated the full left payload even when the join output was empty (COUNT(*)) and no residual filter needed those columns.

This branch keeps the fast take-based output path, but trims the concatenated build batch to just the columns that are actually read later. It also removes two layers of probe-side overhead that were still scalar and allocation-heavy.

Benchmark

Release benchmark:

cargo run --release -p datafusion-bio-function-ranges --example interval_join_bench

bucketed_sparse_count EXPLAIN ANALYZE, IntervalJoinExec with Algorithm::Coitrees:

• base branch perf/issue-33-interval-join: build_time=4.124417ms, join_time=2.035374ms
• this branch: build_time=3.268291ms, join_time=1.883876ms

Representative wall-clock results on the same synthetic dataset (49,152 x 49,152 rows):

• bucketed_sparse_count: 6ms -> 6ms (within noise, but lower operator metrics)
• bucketed_sparse_wide: 8ms -> 8ms (no regression on wide output)
• contig_only_sparse_count: 5ms -> 5ms

The main win is inside the operator metrics for count-style overlap joins, where the build side no longer concatenates unused left payload columns.

Validation

• cargo test -p datafusion-bio-function-ranges
• cargo run --release -p datafusion-bio-function-ranges --example interval_join_bench