Fixes #6014.

Proposes changing the behavior of Config::SetVerbosity() when it's passed params that don't contain verbose or verbosity. Instead of resetting the log level to INFO in those situations, as of this PR it would not modify the log level at all.

See #6014 (comment) for an explanation of this.

Notes for Reviewers

How I tested this

Using the examples in #6014.

And I've added 2 new unit tests here (one for lgb.train(), one for the scikit-learn interface) testing the behavior. These tests fail on master.

I also ran each of the Python test files one at a time, to be sure they weren't accidentally relying on log level state set by previous tests.

pytest tests/python_package_test/test_basic.py
pytest tests/python_package_test/test_callback.py
# etc., etc.

A note on usability

Prior to this change, from the perspective of the Python package it was sort of a hidden implementation detail that the logging level is a process-global variable.

Running lgb.train(dtrain) with all parameters at their defaults would always generate INFO-level logs.

As of this change, passing "verbosity" through params will now more obviously be a way to manipulate global state. For example:

# before: WARN-level logs
# now:     WARN-level logs
lgb.train(dtrain, params={"verbosity": -1})

# before: INFO-level logs
# now:     WARN-level logs
lgb.train(dtrain)

This might confuse some people at first, but I think it's a better behavior .The most important thing to know: as of this PR, passing verbosity through params will always correctly set the log level.