Add references for bandwidth performance

[aurianer]

Implement a performance check on the reference bandwidth, as suggested in #429 (comment).
For now it's testing if 40% of the reference bandwidth is reached, because with 60% some tests were occasionally failing when running manually on beverin. This threshold can be increased in the future.

reframe -C config/cscs.py  -c checks/microbenchmarks/gpu/gpu_benchmarks/parallel_algos.py --system beverin:mi300 -r -v --keep-stage-files

[aurianer]

cscs-ci run alps-beverin-uenv;MY_UENV=prgenv-gnu/25.07-6.3.3:v6

[Copilot]

Pull Request Overview

This PR adds bandwidth performance validation against reference values for GPU parallel algorithms (scan, reduce, and radix-sort). The implementation establishes a baseline requiring tests to achieve at least 40% of reference bandwidth values, with architecture-specific thresholds for mi200, mi300, and gh200 GPUs.

Key Changes

• Added reference bandwidth dictionary containing expected performance values for different algorithms, architectures, and data sizes
• Implemented set_references method to dynamically set performance thresholds based on current partition architecture
• Imported uarch utility to determine GPU architecture at runtime

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jgphpc]

also tested with UENV=prgenv-gnu/25.6:v2 on gh200 (+ UENV=prgenv-gnu/25.07-6.3.3:v6 on mi200/mi300)