`RNN::states_to_tensor` behavior

Currently [`RNN::states_to_tensor`](https://docs.rs/candle-nn/0.3.3/candle_nn/rnn/trait.RNN.html#tymethod.states_to_tensor) implementations for `LSTM` and `GRU` return 2-dimensional tensor of shape `(batch_size, seq_len * hidden_dim)`.

Is this a correct behavior? Shouldn't they return 3-dimensional tensor of shape `(batch_size, seq_len, hidden_dim)` instead?

Start with the linked RNN::states_to_tensor documentation and inspect its LSTM and GRU implementations. Compare their current 2D output with the proposed (batch_size, seq_len, hidden_dim) shape. Done means the expected shape is resolved and the implementation and documentation consistently reflect that decision.

Written by the indexing model from the issue text.