feat(bench): add map-reduce chunking to NIAH benchmark

[starpit]

Summary

Implement chunk-based map-reduce structure in niah.rs benchmark, matching the pattern from haystack.rs. This allows testing retrieval performance with different chunking strategies.

Changes

• Add chunk parameter to run_niah_test() function
• Implement map-reduce logic: split context into chunks, process each chunk separately (map), then combine results (reduce)
• Add BENCH_CHUNK_SIZES environment variable (default: 0,2,4)
• Always include chunk size in benchmark IDs for proper filtering
• Update documentation with chunking examples and filtering guide

Usage Examples

# Run with defaults: chunk=0, chunk=2, and chunk=4
cargo bench --bench niah --features tok

# Filter to run only chunk=0 (non-chunked) benchmarks
cargo bench --bench niah --features tok -- "chunk=0"

# Filter to run only chunk=2 benchmarks
cargo bench --bench niah --features tok -- "chunk=2"

# Run only chunked benchmarks (chunk > 0)
cargo bench --bench niah --features tok -- "chunk=(2|4)"

Made with Bob