[WIP] Add build-flow benchmark tests for FINN example models

[auphelia]

Adds a tests/benchmark/ suite that runs the end-to-end FINN build flow for the public FINN example models across their target boards, checking that all expected build products and per-step verification outputs are produced. This gives us regression coverage for the example designs that today only live in finn-examples.

Also includes one small correctness fix in StreamingDataflowPartition execution so the partition output is cast to the ONNX-declared output datatype during SDP execution.

What's included

• New benchmark tests under tests/benchmark// for:
  • bnn-pynq (CNV/TFC, w1a1/w1a2/w2a2)
  • cybersecurity-mlp
  • gtsrb
  • kws
  • mobilenet_v1
  • resnet50
  • vgg10-radioml
• Each model ships per-board folding_config/ and specialize_layers_config/ JSONs. Boards covered: Pynq-Z1, AUP-ZU3_8GB, ZCU102, ZCU104, U250.
• Each test runs build_dataflow_cfg (generating estimate reports, stitched IP, PYNQ driver, bitfile, deployment package, rtlsim performance) and asserts the expected outputs plus the verify_*_SUCCESS.npy verification artifacts
  (initial/streamlined python, folded HLS cppsim, node-by-node rtlsim, stitched-IP rtlsim).
• tests/benchmark/models/download_models.sh pulls the ONNX models from the finn-examples GitHub releases.
• New finn_examples pytest marker (setup.cfg); .gitignore entries for downloaded models and verification .npy files.

Bug fix

• streamingdataflowpartition.py: propagate the correct output datatype during SDP execution by casting ret[model_oname] to the graph output's declared dtype (previously the raw array dtype leaked through).

Test plan

• pytest -m finn_examples with models downloaded via download_models.sh
• Note Vivado version constraints baked into tests (2022.2 default; 2024.1 required for AUP-ZU3_8GB)

[auphelia]

Currently quicktest is failing, this will pass when #1637 is merged into this branch.