Train a CNN with native CPU replicas

// Native CPU replicas share one global batch and one canonical optimizer update.
import aner.nn
import aner.tensor
import aner.parallel
import aner.dev

let model = nn.cnn(1, 8, 8).conv2d(outputs: 4, padding: 1)
    .batch_norm().relu().head(classes: 2)
    .freeze_batch_norm().on("cpu")
let images = tensor.reshape(nn.random(12, 64, seed: 2026, scale: 1.0), List(12, 1, 8, 8))
let targets = nn.zeros(12, 1)
dev.capture("Parallel CNN architecture", model)
let fitted = model.fit(images, targets, epochs: 3, batch: 5, rate: 0.001,
    execution: parallel.local(workers: 2))
print(fitted.updates()) // 9 updates: every final partial global batch is retained.
print(fitted.accuracy(images, targets)) // API demonstration; these are training inputs.