[R-package] [python-package] deprecate Dataset arguments to cv() and train()

[jameslamb]

Contributes to #6435.

Adds deprecation warnings for all the cases mentioned in #6435, and some other changes that flow from that:

• promotes categorical_feature and colnames to top-level arguments of lightgbm() in the R package
  • so it stops relying on lgb.train()'s support for setting those on the Dataset
• modifies examples and tests to no longer use the deprecated pattern
• updates relevant docs

Notes for Reviewers

I'd like to get these warnings out in the v4.4.0 release (#6439), and then leave them there for at least 2 more releases.

copying some others who may be interested: @mayer79 @david-cortes @simonpcouch

How I tested this

Ran the R tests and Python tests, and checked for these warnings.

# R
sh build-cran-package.sh --no-build-vignettes
R CMD INSTALL ./lightgbm_4.3.0.99.tar.gz
cd R-package/tests
Rscript testthat.R
cd ../../

# python
cmake -B build -S .
cmake --build build --target _lightgbm -j4
sh build-python.sh install --precompile
pytest tests/python_package_test

Interactively ran some code in both languages and confirmed that the expected warnings were raised.

[borchero]

Nice, thanks a lot!

[jameslamb]

@jmoralez what do you think about this?

[jmoralez]

Hey. Sorry, I had given this a thumbs up but I guess my internet decided against it. I'm in favor of this, it'll reduce the complexity of the code and also avoid confusion on which place has the highest priority, etc.

[jameslamb]

Great, thank you! I'm very glad someone asked about this on Stack Overflow... I've been staring at this code for 7 years and never noticed the duplication 😅