12. Backprop: Single Linear Layer

Implement the backward pass for a single linear layer y = xW + b using the chain rule.

Signature: def backprop_single_layer(x: np.ndarray, W: np.ndarray, delta: np.ndarray) -> tuple

Inputs:

• x: input activations of shape (batch, in_features) (or any (..., in_features) for higher-rank inputs)
• W: weight matrix of shape (in_features, out_features)
• delta: upstream gradient dL/dy of shape matching the layer's output (..., out_features)

Returns the tuple (dW, db, dx) of gradients, with shapes:

• dW: same shape as W — (in_features, out_features)
• db: same shape as the bias — (out_features,)
• dx: same shape as x

The bias b itself is not passed in (its gradient depends only on delta). Your implementation should handle both the standard 2D batch (B, in) and higher-rank inputs like (B, T, in) — the test suite includes a 3D case.