Fix gpu metrics

[alanwguo]

cherrypick #56009

Why are these changes needed?

Bugs introduced in #52102

Two bugs:

• proc is a TypedDict so it needs to be fetched via proc["pid"] instead of proc.pid.
• Changing processes_pid is backwards-incompatible change that ends up changing the dashboard APIs that power the ray dashboard. Maintain backwards-compatibility

Verified fix:
Metrics work again:

Ray Dashboard works again:

Related issue number

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[gemini-code-assist]

Code Review

This pull request primarily addresses a bug in GPU metrics reporting by correcting how process IDs are accessed. It also includes a significant refactoring of ray.train.v2 to simplify dataset handling by removing the DatasetManager actor, which improves maintainability. Additionally, it updates the project version to 2.49.0 across various files, including Python, Java, and C++ components, and adds protobuf compatibility helpers for MessageToJson. The tests have been updated to reflect these changes, including a refactoring of aggregator agent tests for better structure. My feedback is minor, pointing out a small typo in a comment.

[can-anyscale]

dang this type of bug should have been caught with a type checker (if we can typify this and other classes)

[alanwguo]

I think TypedDict actually breaks the typechecker.

Autocomplete suggests you can use .field but in reality you cannot.