Enable unsupervised speech dataloading

**Is your feature request related to a problem? Please describe:**
Currently the `GenericSpeechDataset` is [not implemented](https://github.com/facebookresearch/fairseq2/blob/main/src/fairseq2/datasets/speech.py#L76-L103).

**Describe the solution you would like:**
Populate this class to be able to load unlabelled speech corpora such as librispeech

**Describe the alternatives you have considered:**
--

**Additional Context:**
--

Start by reading src/fairseq2/datasets/speech.py around GenericSpeechDataset at lines 76-103, where the issue says the class is not implemented. Trace the existing speech dataset interfaces and determine how unlabelled corpora such as LibriSpeech should be loaded; done means GenericSpeechDataset can load an unsupervised speech corpus.

Written by the indexing model from the issue text.