CI for Benchmarking

[fpjentzsch]

Adds benchmarking suite to GitLab CI, next to the existing test suite.

There will be at least 3 types of benchmark suite:

• "Reduced": A few variants of the most important models, used for regression testing alongside test suite for PRs to dev
• "Full": Large gridsearch intended for fitting QoR estimation models on the result, only run on demand
• Manual: A config file can be submitted via API/Web to run a one-off experiment

Merge branches:
There are currently 2 merge branches to enable 2 types of design-under-test (DUT) that are defined in this PR, but cannot run without additional branches merged in:

• ci_merge_transformer: for Dummy Transformer, GPT Transformer, RadioML Transformer, RadioML ConvFormer
  • transformer merge branch
  • Workaround for MoveOpPastFork: 4204c36
• ci_merge_fifo: for FIFO-sizing-related test of synthetic models, MetaFI, ResNet-50
  • Characterization-based FIFO sizing for non-linear models Xilinx/finn#1229, which is based on Analytical FIFO sizing Xilinx/finn#1185
  • feature/pyxsi_preview, which is based on Pyxsi enablement for alternative rtl simulation Xilinx/finn#1213
  • Xilinx/finn/feature/fifo_gauge_integration for infinite queue-based FIFOs during simulation
  • Update QONNX to current main
  • Make N for largefifo_rtlsim configurable: fe22a5e
  • Make throttling for largefifo_rtlsim configurable: 2572ecf

TODO:

• Benchmark suite selector
• Switch from JSON to YAML?
• Upload of key metrics to central location for monitoring of regression tests
• Parse more metrics from the build flow
• Cleanup and refactoring

Extensions:

• Integrate new instrumentation wrapper/harness
• Integrate new measurement setup

DUTs to add:

• MVAU
• VVAU
• SWG
• Thresholding
• Synthetic non-linear model (for testing FIFO-sizing)
• Transformer: Dummy, GPT, RadioML, RadioML ConvFormer
• ResNet-50
• MobileNetV1
• VGG-10
• MetaFi