HF Model loading fixes

[farzadab]

This PR fixes the following:

1. gets rid of the unexpected weights warnings
2. makes HF model loading faster by skipping random initialization

Example of the unexpected weights warning that this PR is supposed to resolve:

Some weights of the model checkpoint at fixie-ai/ultravox_dev were not used when initializing UltravoxModel: ['language_model.lm_head.weight', 'language_model.model.embed_tokens.weight', 'language_model.model.layers.0.input_layernorm.weight',

...

- This IS expected if you are initializing UltravoxModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing UltravoxModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).