fix: correct Bi-LSTM output shape comment (n_hidden*2 not n_hidden)

[Mukller]

Fix: Incorrect output shape comment in Bi-LSTM(Attention)

The comment for the LSTM output after permute(1, 0, 2) incorrectly states
[batch_size, len_seq, n_hidden] when it should be [batch_size, len_seq, n_hidden * num_directions(=2)].

A bidirectional LSTM concatenates forward and backward hidden states, so the
last dimension is n_hidden * 2, not n_hidden.

Before (wrong comment):

output = output.permute(1, 0, 2) # output : [batch_size, len_seq, n_hidden]

After (correct comment):

output = output.permute(1, 0, 2) # output : [batch_size, len_seq, n_hidden * num_directions(=2)]

This matches the existing comments on surrounding lines (e.g., attention_net's
docstring already correctly states n_hidden * num_directions(=2)).

Closes #84