Add generic dataset

[zqhuang211]

Add GenericVoiceDataset to support new datasets via yaml config, e.g.,

data_dicts:
  - path: "fixie-ai/librispeech_asr"
    name: "clean"
    splits:
      - "train.100"
      - "train.360"
    user_template: "Continue the following text using less than 50 words:\n\n<|audio|>"
    assistant_template: "{{ continuation }}"
    transcript_template: "{{ text }}"
    num_samples: 100_000
  - path: "fixie-ai/librispeech_asr"
    name: "other"
    splits:
      - "train.500"
    user_template: "Continue the following text using less than 50 words:\n\n<|audio|>"
    assistant_template: "{{ continuation }}"
    transcript_template: "{{ text }}"
    num_samples: 100_000

See ultravox.training.configs.llama3_whisper_kd.yaml for example experiment configuration.

[zqhuang211]

The previous PR was not merged properly. Submitting it one more time.