move normalize and denormalize into perturb

[yairchn]

Earth-2 MIP Pull Request

Description

This PR move the normalize and denormalize into perturb as a first step to decouple the normalization (which is model dependent) from the time loop. To do this requires providing perturb with the means and stds of the channels for all models. To that I added two numpy files to channel_means.py and channel_stds.py which are dicts that contain the largest dataset we currently have (235 channels).

Checklist

• [x] I am familiar with the Contributing Guidelines.
• [x] New or existing tests cover these changes.
• [x] The documentation is up to date with these changes.
• [x] The CHANGELOG.md is up to date with these changes.
• [x] An issue is linked to this pull request.

Dependencies

[yairchn]

I have tested a small ensemble inference in the notebook and the results below look as expected

[nbren12]

/blossom-ci

[yairchn]

/blossom-ci

[nbren12]

/blossom-ci

[nbren12]

/blossom-ci

[NickGeneva]

/blossom-ci

[nbren12]

There is a test failure looking up the missing "a" data: https://github.com/NVIDIA/earth2mip/actions/runs/6475309470/job/17582074381#step:2:923

Suggest manually adding it to the means/stds dicts in the test function. Or you could change the default behavior of normalization to e..g. not perturb variables with no included mean/var.

[NickGeneva]

/blossom-ci