Which issue does this PR close?

N/A - new tooling

Rationale for this change

The existing shuffle benchmarks use small synthetic data (8192 rows x 10 batches) with Criterion, which makes it difficult to:

• Benchmark with realistic data distributions from TPC-H/TPC-DS at scale
• Profile with tools like cargo flamegraph, perf, or instruments (Criterion's harness interferes)
• Test both write and read paths (current benchmarks are write-only)
• Explore different scenarios like spilling, high partition counts, or codec comparisons

What changes are included in this PR?

Adds a shuffle_bench binary (native/core/src/bin/shuffle_bench.rs) that benchmarks Comet shuffle write and read performance independently from Spark.

Features:

• Parquet input: Point at TPC-H/TPC-DS Parquet files for realistic data distributions
• Synthetic data generation: Configurable schema with int, string, decimal, and date columns
• Write + read benchmarking: --read-back decodes all partitions and reports throughput
• Configurable scenarios: partitioning (hash/single/round-robin), partition count, compression (none/lz4/zstd/snappy), memory limit for spilling
• Profiler-friendly: Single long-running process with warmup and iteration support

Example usage:

# Benchmark with TPC-H data
cargo run --release --bin shuffle_bench -- \
  --input /data/tpch-sf100/lineitem/ \
  --partitions 200 --codec zstd --read-back

# Generate synthetic data
cargo run --release --bin shuffle_bench -- \
  --generate --gen-rows 10000000 \
  --partitions 200 --codec lz4 --read-back --iterations 3 --warmup 1

# Profile with flamegraph
cargo flamegraph --release --bin shuffle_bench -- \
  --input /data/lineitem/ --partitions 200 --codec zstd

How are these changes tested?

Manually tested with both generated data and various configurations:

• Different codecs (none, lz4, zstd, snappy)
• Read-back verification (all rows decoded correctly)
• Multiple iterations with warmup
• Clippy clean, cargo fmt applied