don't resize embeddings if it's already large enough

[winglian]

for example, we don't actually need to resize phi b/c the embeddings size is already much larger than the tokenizer len

>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-1_5", trust_remote_code=True)
>>> len(tokenizer)
50295
>>> model = AutoModelForCausalLM.from_pretrained("microsoft/phi-1_5", trust_remote_code=True)
>>> model.get_input_embeddings().num_embeddings
51200
>>> 

[NanoCode012]

As I rmb, model.resize implicitly calls model.tie_weights. By doing this, we're not tying weights. I'm not clear whether this is still important as it was a warning message while finetuning a month ago.

It is also weird that the tokenizer's length and the model's weights have different length..