WIP: [python-package] allow access to eval result attributes by name

[jameslamb]

Closes #6748

Wraps evaluation result tuples in the Python package in a typing.NamedTuple, to allow named-attribute access to eval result tuples.

NOTE: If you see this, this isn't ready for review. Just opening to show I'm working on it and get feedback from CI.

Notes for Reviewers

Benefits of this change

• simpler, less error-prone code
• improves readability
• improves the ability of type-checkers like mypy to catch bugs
• eliminates awkward workarounds for different-sized tuples with and without a standard deviation (only added by cv())
• tab completion when modifying the code

And provides all of these benefits for code that uses lightgbm, especially code that touches evaluation results like optuna.

Is this safe?

I think so. Any existing code treating these values (like the output of Booster.eval_valid()) as Python tuples should continue to work exactly as they always have.

References

See https://snarky.ca/dont-use-named-tuples-in-new-apis/