[ci] test 32-bit R in CI

[jameslamb]

Based on #3585 (comment). We were recently surprised by some test failures on 32-bit Windows checks for CRAN.

I think this is because I thought we were running tests with a 32-bit version of R on Windows (#3311 ) but maybe we never actually were.

I think that the root cause is that the test setup for our R environment doesn't install i386 components for R-win.exe, as documented in https://cran.r-project.org/bin/windows/base/rw-FAQ.html#Can-I-customize-the-installation_003f.

As of this PR, 32-bit R will be installed in GitHub Actions jobs and the R package will be tested against both architectures (x86 and i386).

Questions for Reviewers

In #3585 , we ran into issues only on CRAN's R 3.6 + Windows (32-bit) job. Today, the only tests of the CRAN package are on R 4.0. That's a gap between this project's CI and CRAN, which means we could run into this situation again in future CRAN submission. In this PR I'd like to also add an R 3.6 + Windows + build_type: cran job.

GitHub Actions gives you 20 concurrent job runs, and right now we have 17 jobs that run on every PR.

How do you feel about adding this additional job?

          - os: windows-latest
            task: r-package
            compiler: MINGW
            toolchain: MINGW
            r_version: 3.6
            build_type: cran

NOTE: this PR's tests won't pass until #3586 is merged. I wanted to keep them separate so this CI work doesn't slow down the process of releasing a fix for CRAN.

[jameslamb]

I think this worked! Now the behavior in CI matches what we're seeing on CRAN...the R3.6 + cran build is failing with the exact same issues described in #3585

https://github.com/microsoft/LightGBM/pull/3588/checks?check_run_id=1437170140

  == testthat results  ===========================================================

  FAILURE (test_learning_to_rank.R:52:9): learning-to-rank with lgb.train() works as expected

  FAILURE (test_learning_to_rank.R:53:9): learning-to-rank with lgb.train() works as expected

  FAILURE (test_learning_to_rank.R:130:5): learning-to-rank with lgb.cv() works as expected

  FAILURE (test_learning_to_rank.R:136:5): learning-to-rank with lgb.cv() works as expected

And also a new compiler warning we haven't noticed before

./include/LightGBM/utils/common.h:47:0: warning: "_mm_free" redefined
 #define _mm_free(a) free(a)
 ^

./include/LightGBM/utils/common.h:46:0: warning: "_mm_malloc" redefined
 #define _mm_malloc(a, b) malloc(a)
 ^

[StrikerRUS]

How do you feel about adding this additional job?

Totally fine!

[jameslamb]

@StrikerRUS I just moved the changes to the tests from #3586 over to this PR. Since it seems like #3586 might be delayed for an unknown amount of time waiting for CRAN to reply, I think we should merge this PR with the updated tests + the new CI job sooner.

I want this new CI job to be sure that new bugfixes and features (for example, #3405) don't break the 32-bit CRAN checks.

[jameslamb]

checked the logs one more time and I can see the tests are now running for both i386 and x64 architectures

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.