[python-package] factor out positional indexing for eval result tuples

[jameslamb]

Description

While working on #3756 and #3867, as well as a variety of early-stopping related things (like #6424), I've found the heavy use of positional indexing and tuple unpacking in early stopping and related code make it really difficult to understand and modify.

For example:

LightGBM/python-package/lightgbm/callback.py

Lines 167 to 176 in 53e0ddf

	for item in env.evaluation_result_list:
	if len(item) == 4:
	data_name, eval_name, result = item[:3]
	self.eval_result[data_name][eval_name].append(result)
	else:
	data_name, eval_name = item[1].split()
	res_mean = item[2]
	res_stdv = item[4] # type: ignore[misc]
	self.eval_result[data_name][f"{eval_name}-mean"].append(res_mean)
	self.eval_result[data_name][f"{eval_name}-stdv"].append(res_stdv)

and:

LightGBM/python-package/lightgbm/callback.py

Lines 410 to 412 in 53e0ddf

	score = env.evaluation_result_list[i][2]
	if first_time_updating_best_score_list or self.cmp_op[i](score, self.best_score[i]):
	self.best_score[i] = score

I'm opening this issue to track some work I'd like to do to simplify that.

Benefits of this work

Reduces the effort required to finish these:

• [python-package] type hints in python package #3756
• [ci] warnings from mypy #3867

And to add finer-grained control over early stopping and validation, e.g.:

• [WIP] Add chosen metric argument to clarify early stopping behaviour #6424 (comment)

Approach

I'm planning a series of PRs with the following types of changes:

• unpacking tuples into named variables and using those named variables
• introducing collections.namedtuple (docs) where appropriate to allow for named-attribute access while preserving backwards compatibility with the significant amount of custom code in the world relying on lightgbm expecting tuples for this like custom metrics:

LightGBM/python-package/lightgbm/engine.py

Lines 135 to 138 in 53e0ddf

	feval : callable, list of callable, or None, optional (default=None)
	Customized evaluation function.
	Each evaluation function should accept two parameters: preds, eval_data,
	and return (eval_name, eval_result, is_higher_better) or list of such tuples.

Notes

Looking at the current state of this code, it's important to remember that when the lightgbm Python package was first introduced 8+ years ago:

• it supported Python 2.7
• dataclasses was not in the standard library yet (that came in Python 3.7, first release in June 2018)