fix: CSPLayerWitnTwoConv originally involves dynamic forwarding, whic…

[MR-hyj]

Motivation

During my QAT training with mmrazor and mmyolo, it reported that some original designs of mmyolo layers can't be traced by troch.fx, because they work in a dynamic way, (e.g. the use of list).

So I changed the way it forwards, without any use of list, dict, or tuple, running with tensor instead.

Modification

change the calculation by avoiding any usage of list

BC-breaking (Optional)

No

[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.

---

hengyujiang 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.}