Added fixed-size sequence length for GRU

[ColdFrenzy]

When using an RNN, the sequence length was directly linked to n_envs_per_worker:

sequence_length = -(
    -self.experiment_config.collected_frames_per_batch(self.on_policy)
    // self.experiment_config.n_envs_per_worker(self.on_policy)
)

For large values of collected_frames_per_batch, this requires a high number of n_envs_per_workers. This can cause problems or overhead in non-vectorized and CPU environments.
The goal of this pull request is to remove this dependency and use fixed sized sequence length by leveraging the hidden state obtained during the collection phase.