Explore GPU-accelerated query execution via Sirius/CUDA

## Context

[Sirius](https://www.sirius-db.com/) is an open-source (Apache 2.0) DuckDB extension by NVIDIA + University of Wisconsin-Madison that adds GPU-accelerated query execution via CUDA/cuDF. It [set ClickBench records in January 2026](https://developer.nvidia.com/blog/nvidia-gpu-accelerated-sirius-achieves-record-setting-clickbench-record/), achieving at least 7.2x higher cost-efficiency vs CPU.

**How it works:** DuckDB handles parsing/optimization/scanning on CPU → data converts to Arrow → cuDF executes operators on GPU (joins, aggregations, projections, regex) → results back to CPU.

## Why Not Today

- **Go bindings:** Arc uses DuckDB via Go. Loading a CUDA extension requires NVIDIA runtime on every deployment.
- **NVIDIA-only:** Requires NVIDIA GPUs. Most Arc deployments are CPU-only VMs/containers.
- **Immature:** Single-node, single-GPU. No GPU-native Parquet reading. Some operators fall back to CPU. No disk spilling.
- **Arc's bottleneck is I/O, not compute:** Reading Parquet from disk/S3 and serializing results dominates query time. GPU helps most on heavy aggregations over data already in memory.

## When to Revisit

- Sirius supports multi-GPU, disk spilling, and GPU-native Parquet reading
- Arc adds a hot cache layer (data kept in GPU memory across queries)
- Customer demand for heavy analytical workloads on GPU-equipped infrastructure
- Potential "Arc Enterprise Analytics" tier on GPU hardware

## References

- [Sirius GitHub](https://github.com/sirius-db/sirius)
- [NVIDIA blog post](https://developer.nvidia.com/blog/nvidia-gpu-accelerated-sirius-achieves-record-setting-clickbench-record/)
- [DuckDB GPU discussion](https://github.com/duckdb/duckdb/discussions/8675)