Expose last hidden state on RNNModel and BlockRNNModel

[AmjadAAYD]

Fixes #1649. Both models already compute a hidden state internally but discarded it. Now stashed as an attribute on the underlying PyTorch module during forward/predict, and exposed via a last_hidden_state property on both RNNModel and BlockRNNModel (None before first predict/fit). For LSTM this is the (h_n, c_n) tuple, for RNN/GRU a single h_n tensor, matching what the underlying torch module returns. Verified with a real fit+predict on both LSTM and GRU variants of both models, checking shapes and types.

[AmjadAAYD]

Checked everything, tests pass locally. CI is waiting on a maintainer to approve the workflow run, standard for a first PR from a new contributor.

[AmjadAAYD]

Checked everything on my end, tests pass locally. CI hasn't run yet, it needs a maintainer to approve the workflow run since this is my first PR to this repo.