ReLU & GELU Activations

Activation functions add the non-linearity that lets neural networks learn complex patterns.

Implement three vectorised functions:

• relu(x): $\max(0, x)$
• relu_grad(x): the derivative of ReLU, which is 1.0 where $x > 0$ and 0.0 elsewhere
• gelu(x): the tanh approximation used by GPT-2 and BERT:

All three take and return NumPy arrays of any shape.