Follow-up to #750 (comment)

• Return all beams (tensor of shape (batch_size, num_beams, sequence_length)).
• Return the beam scores as well (tensor of shape (batch_size, num_beams)).

The changes are to be made to this utility function: https://github.com/keras-team/keras-nlp/blob/master/keras_nlp/utils/text_generation.py#L213.

@mattdangerw, do we want to sort the outputs based on log prob and return? Or is that unnecessary overhead?

Some basic tinkering done because a user wanted a quick fix: https://github.com/abheesht17/keras-nlp/blob/beam_search_return_all_seqs/keras_nlp/utils/text_generation.py#L400-L414