[R-package] [ci] restore R 3.6 Windows cran CI job (fixes #5036)

[jameslamb]

Fixes #5036.
Replaces #5037.

Changes in this PR

• changes to build-cran-package.sh:
  • replaces use of tar --exclude with simply rm-ing the excluded files prior to re-tarring the source distribution
  • removes unnecessary uses of -f in rm commands deleting files (not directories), for safety
  • removes unnecessary use of -v in tar commands whose output is suppressed anyway by redirection to /dev/null
  • removes uses of tar -C, instead just relying on directory traversal with cd
• restores the R 3.6 cran-style Windows CI job

Description

CRAN-style builds of the R package have not been tested against R 3.6 on Windows for the last 7 months. This PR restores the CI job testing that support.

I used #5037 to experiment with different approaches and to try to find a root cause for why this job started failing. I discovered that the root issue is related to conflicting versions of the tar utility. As described in the R documentation for ?tar (link)

The ‘tar’ format no longer has an agreed standard!

Many R platforms use a version of GNU tar (including Rtools on Windows), but the behaviour seems to be changed with each version.

This only matters to {lightgbm} because its build script, build-cran-package.sh, requires a workaround for R not cleaning up .o files produced when the library is compiled to build vignettes (see #3946 (review)). That workaround requires untarring and carefully re-tarring a source distribution built by R CMD build.

This PR removes uses of less-likely-to-be-portable tar features in that build script, like -C and --exclude, to make failures like the one from #5036 less likely...in CI and for LightGBM's users.

[StrikerRUS]

Thanks a lot! Great!

[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.