Implement a function that applies dropout to a matrix of attention weights (the output of softmax in scaled dot-product attention) and then computes the resulting context vectors by multiplying with the value matrix.
Given:
attn_weights: a 2D numpy array of shape (T, T) of post-softmax attention weightsvalues: a 2D numpy array of shape (T, d_v) of value vectorsdropout_rate: a float in [0, 1) giving the probability of zeroing each entrymask: a 2D numpy array of 0s and 1s with the same shape as attn_weights. Entries where mask == 0 are dropped (set to zero); surviving entries are scaled by 1/(1 - dropout_rate)Return the context matrix of shape (T, d_v) as a nested Python list (via .tolist()).
If dropout_rate is 0, return values unchanged by dropout (no scaling, no zeroing).