Second moment statistics based on dataset

[JakubResh]

Implement the second moment statistics matrix computation using the wikipedia dataset

[JakubResh]

Possibly not needed - in latest versions, the random approach seems to work well enough on the same amount of sequences (100 000 random token sequences)

[JakubResh]

We need more experiments. So far, the random-based second moment seems like a cause for hallucinations. Once we implement more capable models, we can test this

[olexamatej]

https://github.com/zjunlp/EasyEdit/blob/main/easyeditor/models/rome/layer_stats.py
https://github.com/qizhou000/OpenEdit
https://github.com/ChenhuiHu/WilKE

some inspiration i found

[olexamatej]

figured it out!, I will push it to branch made for MSD after I will be done with changes, you can then cherrypick the commit

1 issue I came upon is that we cannot use float16, and we NEED float32

[olexamatej]

commit #8fe83 on MSD branch has the full implementation of second_moment_wikipedia

1. I couldnt make it work with float16, so I have changed multiple parts of codebase to float32
2. I had to add some normalization here and there
3. There are new config parameters, one is for batch vs interactive mode (interactive just runs quick demo, batch mode uses counterfact and tests it better)
4. I moved some parts from insert_kv to my second_moment_wikipedia, so the random one probably wont work on this branch

Its just a demo, we need to make sure it works, then I can cherrypick it onto different branch from MSD