Implement a function that takes a square matrix of attention weights (already normalized by softmax) and returns a causally-masked, re-normalized version using two equivalent approaches.

Given attention weights of shape (T, T) and a string method:

• If method == 'tril': construct a lower-triangular mask of 1s (including diagonal) of shape (T, T), multiply elementwise with the attention weights, then re-normalize each row so it sums to 1.
• If method == 'triu': construct an upper-triangular boolean mask above the diagonal (k=1) of shape (T, T); using the ORIGINAL pre-softmax-like scores reconstructed by taking log of the attention weights, set masked positions to -inf, then apply softmax along the last axis.

Both methods must produce the same final matrix (up to floating point). Return the resulting (T, T) matrix as a nested Python list.

Assume the input attention weights are strictly positive (so log is defined) and each row sums to 1.

Parameters:

• attn_weights: numpy array of shape (T, T), each row sums to 1
• method: 'tril' or 'triu'

Return: nested list (T, T) of the causally masked, normalized attention weights.