Add resnet20

[stamcenter]

Adding the second pull request which contains the resnet-20 model as the reference for cifar10. It also include a modification to the harness that allow it to set mlp as default model for mnist and resnet20 as default model for cifar10.

[code-perspective]

Nice! I was able to run this.

Can you rebase against main? Please only keep the cifar-10 and resnet20 models in this PR.

For reference, I got the measurement files (roughly takes 2 hours to run with the cifar-10 dataset load time).

{
  "model_name": "resnet20",
  "dataset_name": "cifar10",
  "Timing": {
    "Test dataset generation": "28.5534s",
    "Key Generation": "167.453s",
    "Encrypted model preprocessing": "0.0102s",
    "Input generation": "6.0968s",
    "Input preprocessing": "0.0073s",
    "Input encryption": "1.4808s",
    "Encrypted computation": "5256.1713s",
    "Result decryption": "1.0469s",
    "Result postprocessing": "0.0072s",
    "Total": "5460.8268s"
  },
  "Bandwidth": {
    "Public and evaluation keys": "50.4G",
    "Encrypted input": "28.0M",
    "Encrypted results": "8.0M"
  },
  "Server Reported": {}
}

[code-perspective]

In C++20 standard libraries, std::unique is marked [[nodiscard]], which triggers compiler errors when the return value is ignored. Additionally, std::unique and std::remove only reorder elements without actually resizing the underlying vector.

Let's clean this up using standard erase-remove and erase-unique idioms

std::sort(rotation_positions.begin(), rotation_positions.end());
        rotation_positions.erase(std::unique(rotation_positions.begin(), rotation_positions.end()), rotation_positions.end());
        rotation_positions.erase(std::remove(rotation_positions.begin(), rotation_positions.end(), 0), rotation_positions.end());

[code-perspective]

The code to pack the weights can be added in this file.