Replace arbitrary sequence length kernel wrapper with a padding wrapper.

[maximilianmbeck]

This PR replaces the default wrap_chunkwise__arbitrary_sequence_length kernel wrapper for inference (more precisely for prefill with the wrap_chunkwise__arbitrary_sequence_length_with_padding kernel wrapper.

We keep the old wrapper functions for reference.

In short, the new wrapper relies on the insight that we can pad the qkv sequence from the right with zeros and set the forget gate to 1 and the input gate to 0, so that the final memory state is maintained over the padding time steps.

This is more efficient than the previous arbitrary sequence length kernel wrapper, which relied on a sequence of chunkwise and step kernel calls to precisely match the arbitrary sequence length.

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}

[maximilianmbeck]

I have read the CLA Document and I hereby sign the CLA

[maximilianmbeck]

recheck