replace skip_embed with input_embeds

[TideDra]

This PR replace the arg skip_embed: bool with input_embeds: torch.Tensor in the forward functions of all the language models.

This fixes the bug raised by LogitsProcessor when using llava in Extend mode and return log_prob, such as using gen(choices=[...]) with llava, where LogitsProcessor need to receive input_ids and uses it to get the token logprob. However, the input_ids given to LogitsProcessor by llava is actually the input embedding, which will cause a Tensor dimension mismatch error.

[TideDra]

With this fix, we can use gen(choices=[...]) with llava in text-only mode but can not give images, such as:

@function
def test(s,img_path,question):
    s += user(image(img_path)+question+" Please answer yes or no.")
    s += assistant(gen('answer',choices=['yes','no']))

This will raise Runtime error, and it seems to be caused by the cache system not working well with image input.

[aliencaocao]

Same observation with runtime error for image inputs