[WIP][Bug fix] use pad_shape instead of img_shape for bevformer and recover the origin code where the first frame pre_bev is None

[FlyingQianMM]

• Use pad_shape instead of img_shape for bevformer, where img_shape reserves the origin image shape and pad_shape is the image shape after padded.

• Recover the origin codes where the first frame pre_bev is None. Present codes set the first frame pre_bev to zeros during the trainng phase in order to keep consistent with the deploy phase, but it seems to be not too right. The reasons are two below:

1. Temporal information cannot be closed flexiblely where the pre_bev should be set as None.
2. When temporal information is needed, pre_bev is the key and value of the temporal encoder layers. If pre_bev is None, the encoder layer will set the value of pre_bev the same as query.

[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.

---

luoqianhui 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.}

[nepeplwu]

This PR conflicts with the latest code, please fix it