feat: benchmark suite comparing performance against PCL and Open3D

[rajgandhi1]

Overview

Create a published benchmark suite comparing 3DCrate's performance against PCL (C++) and Open3D (Python/C++) on common robotics operations. Publish results in the README and as a GitHub Pages report.

Motivation

The robotics community is conservative about switching tools. Performance numbers are the strongest adoption argument. Community research shows this is how teams evaluate new libraries — they won't migrate without data.

Operations to Benchmark

Operation	Input size	Metric
Voxel grid downsampling	300K, 1M, 5M pts	ms/cloud
Statistical outlier removal	300K, 1M pts	ms/cloud
Normal estimation (KNN k=10)	300K, 1M pts	ms/cloud
ICP (point-to-point)	50K → 50K pts	ms/iter, final MSE
RANSAC plane segmentation	300K pts	ms/cloud
Euclidean clustering	300K pts	ms/cloud
PLY read (binary)	1M pts	ms, MB/s
PLY write (binary)	1M pts	ms, MB/s

Methodology

• Same input data for all libraries (shared .ply/.pcd files checked into benches/data/)
• 100 warm-up runs, report median ± p95
• Hardware: document CPU/RAM/GPU used
• Single-thread and multi-thread results where applicable
• For Python (Open3D): use timeit with process isolation
• For Rust (3DCrate, PCL via C FFI): use criterion

Output

• benches/README.md with tables and charts
• GitHub Actions job that regenerates results on main pushes
• Badge in top-level README: ![ICP speed](https://...)

Notes

• PCL benchmarks should use their standard CMake build (Release mode, -O3)
• Open3D benchmarks should use their official PyPI wheels
• The threecrate-bench crate already exists — extend it with these cases