Gradient Clipping by Global Norm

A single bad batch can produce a huge gradient that wrecks training. Gradient clipping rescales all gradients together when their combined size is too large.

Implement clip_grad_norm(grads, max_norm):

• grads is a list of NumPy arrays (one per parameter, any shapes).
• The global norm is $\sqrt{\sum_i \|g_i\|^2}$ over every element of every array.
• If it exceeds max_norm, multiply every gradient by max_norm / global_norm. Otherwise leave them as they are.
• Return (clipped_grads, global_norm), where global_norm is the value before clipping. Don't modify the input arrays.