enable correct padding_idx for embedding layers

[gupta-abhay]

• This functionality is primarily needed when you call the resize_position_embeddings call for a HF model (ie, adding new tokens to vocab). If the pad_token already exists for the config, you need to ensure that it gets set correctly post _init_weights call.
• By default, the behavior remains unchanged from today, since the pad_token_id - if not specified by the config, defaults to None from here: https://github.com/huggingface/transformers/blob/174890280b340b89c5bfa092f6b4fb0e2dc2d7fc/src/transformers/configuration_utils.py#L326 and the padding_idx is equivalently set to None

[dakinggg]

can you add a simple test for the new init function?