Implemented Self Attention for Question Encoding

[Drajan]

Multi-head Attention weights are computed using two FF layers with ReLU activation and softmax for probabilities ==> softmax(FF(ReLU(FF(input))))
input = [batch_size, num_words, embed_size]
attention = [batch_size, num_words, num_attention_heads]
input_attention_weighted = [batch_size, num_heads, embed_size]
output = [batch_size, num_heads*embed_size] ==> concatenating multi-head representation