feat(mlflow auth): Support for multiple servers

[gmertes]

Description

Implement storing tokens from multiple mlflow servers on disk, so that the user can switch between them without having to login every time.

For example, the following snippet will log in once to a prod and test server, store both server tokens on disk, and authentication can then happen to both servers without having to log in again:

from anemoi.utils.mlflow.auth import TokenAuth

# log in to prod and test, stores both tokens on disk
TokenAuth("http://prod.server").login()
TokenAuth("http://test.server").login()

# read the respective server tokens from disk and authenticate
TokenAuth("http://prod.server").authenticate()
TokenAuth("http://test.server").authenticate()

Practically for users of anemoi-training: it means your training runs can switch seamlessly between logging to prod and test (assuming you logged in to both servers once before).

Notes

Pydantic is used to validate the data on disk. This file is loaded into a ServerStore, which itself is a RootModel dictionary of ServerConfig objects, where the URL of each server is the root dictionary key. Each ServerConfig contains the token information for a given server.

I am deprecating the load_config function, which used to directly return the underlying data format. That function was there so that downstream code in anemoi-training could get the last used server URL from the config file. Now that there are multiple servers, that is handled by a different abstraction get_servers. There's also no need for the underlying data format to be part of the public API. But for now it is still there in deprecated form, and everything is still backwards compatible.

After merging this PR I will update the downstream code in anemoi-training to stop using load_config, and then we can eventually remove it after some time.

Data format

The legacy data format stored in the json was just a simple dictionary with all information for one server:

{
  "url": "https://server.url",
  "refresh_token": "refresh-token",
  "refresh_expires": 123
}

The format has now changed into a dictionary of servers, where url is no longer a member but used as index:

{
  "https://server-1.url": {
      "refresh_token": "refresh-token-1",
      "refresh_expires": 1,
  },
  "https://server-2.url": {
      "refresh_token": "refresh-token-2",
      "refresh_expires": 2,
  },
}

Going from legacy to new format should be seamless for users, config files will be converted into the new format next time a user logs in to a server. There are tests for both legacy and new format.

What issue or task does this change relate to?

#188

As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/

By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.

[anaprietonem]

left question about the use of this feature with syncing but otherwise LGTM!