106. PyTorch: Cosine Annealing LR

Implement a cosine annealing learning rate schedule using PyTorch's built-in CosineAnnealingLR scheduler.

Other PyTorch scheduler problems (e.g. step LR, warmup, ReduceLROnPlateau) are covered as separate follow-up problems.

Signature: def cosine_lr_schedule(lr_init, T_max, n_steps)

• lr_init: initial learning rate (float)
• T_max: number of steps for half the cosine cycle (int)
• n_steps: total steps to simulate (int)
• Returns: list of learning rates at each step (length = n_steps)

Steps:

1. Create a dummy torch.nn.Parameter and torch.optim.SGD optimizer with lr=lr_init
2. Create torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=T_max)
3. For each step, call scheduler.step() and record optimizer.param_groups[0]['lr']
4. Return the list of recorded learning rates

The cosine annealing formula (with eta_min=0):

lr(t) = 0.5 * lr_init * (1 + cos(π * t / T_max))