[feat] Omit padded tokens in generate() of huggingface transformers

[XbzOnGit]

Currently generate() of huggingface transformers pads all requests in a batch to the same length both in prefill and in decode.
This leads to offloading meaningless padded tokens, which can be a lot if some requests are notably longer than others.

[XbzOnGit]

This also requires changing start_token_idx in submit to host engine. Currently it assumes all the requests in a batch share the same prefix length and has the same effective length.
A clean way would be breaking the batch at GPU side, and submit a list of requests to host engine, each with separate prefix length and effective length.

[XbzOnGit]

Another option: Slice at both sides. Dynamic decisions on offloading some useless tokens then drop at host side(achieve larger throughput in transmission of PCIE) and dropping at GPU side(if a lot of useless tokens)

[XbzOnGit]

Padded tokens are not offloaded to database now(resolved by #42 ). Dropping at GPU side is a potential future work.