Evaluate a chronological sequence workflow

// Split chronological segments before windowing. These scaled synthetic
// values need no fitted transform; fit real scaling only on training data.
import aner.nn

let train_series: Tensor = [
    [-0.9], [-0.8], [-0.7], [-0.6], [-0.5],
    [-0.4], [-0.3], [-0.2], [-0.1], [0.0]]
let valid_series: Tensor = [[0.1], [0.2], [0.3], [0.4], [0.5], [0.6]]
let x = nn.sequence_windows(train_series, window: 3)
let y = nn.sequence_targets(train_series, window: 3)
let valid_x = nn.sequence_windows(valid_series, window: 3)
let valid_y = nn.sequence_targets(valid_series, window: 3)
let model = nn.lstm(inputs: 1, hidden: 8, seed: 17).on("cpu", "Float64")
let fitted = model.fit(x, y, epochs: 12, batch: 4,
    rate: 0.01, clip_norm: 1.0)
let validation = fitted.evaluate(valid_x, valid_y, batch: 2)
print(validation.summary())
print(validation.samples())
print(validation.observations())
print(fitted.diagram())
// train_step keeps optimizer moments across independent loader minibatches;
// fit_chunk instead carries detached memory through ordered streaming chunks.
// fitted.save("sequence-workflow.anerrnn")