Fix freeze_stage in swin.py and bug of split_outputs in dino_head.py

[ShenghaiRong]

Motivation

1. The motivation behind this PR is to fix the error that occurs when training grounding-dino-swin-t with the argument frozen_stage set in a distributed training. The error occurs because swin.py misses adding self._freeze_stages in init(), leading to the optimizer adding the parameters of swin.
2. When not using dn, there would be a TypeError in the split_outputs function in dino_head.py.

Modification

In this PR

1. We add self._freeze_stages in init() in swin.py
2. We modified the split_outputs in dino_head.py

[CLAassistant]

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.

---

Shenghai Rong seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}