[CI] CI cost saving

[hcho3]

• Drop CUDA 10.x and use CUDA 11.0 in CI
• Drop JDK 11 and JDK 12 tests

[hcho3]

I'll need to rebuild the machine image for the Windows worker to use CUDA 11.0.

[hcho3]

I'll keep #7366 open, since clang-tidy fails when using CUDA 11.0

[trivialfis]

Is it caused by def dockerArgs = "--build-arg CUDA_VERSION_ARG=10.1" in the ClangTidy groovy function?

[hcho3]

When I changed 10.1 to 11.0, clang-tidy fails to pass.

[trivialfis]

Got it. I will take a look into the error later.

[trivialfis]

Hi, the errored CI tests for clang-tidy seems to be cmake failing to find appropriate CTK:

[2021-11-16T09:29:31.963Z] -- Detecting CUDA compiler ABI info
[2021-11-16T09:29:32.615Z] -- Detecting CUDA compiler ABI info - done
[2021-11-16T09:29:32.615Z] CMake Error at CMakeLists.txt:134 (message):
[2021-11-16T09:29:32.615Z]   CUDA version must be at least 11.0!
[2021-11-16T09:29:32.615Z] 
[2021-11-16T09:29:32.615Z] 
[2021-11-16T09:29:32.615Z] -- Configuring incomplete, errors occurred!
[2021-11-16T09:29:32.615Z] See also "/workspace/cdb/CMakeFiles/CMakeOutput.log".

[trivialfis]

Other than clang-tidy, is there anything else we need to do for this PR?

[hcho3]

@trivialfis We can merge this PR once the Windows test completes running. I had to re-build the Windows machine's image from scratch.

[hcho3]

@trivialfis Urr... One of the gtest is now failing:

[ RUN      ] MultiClassesSerializationTest.GpuHist
tests/cpp/test_serialization.cc(49): error: Expected equality of these values:
   l
     Which is: false
   r
     Which is: true

[hcho3]

@trivialfis Are you okay with disabling MultiClassesSerializationTest.GpuHist for now? Reducing CI cost is a higher priority.

[trivialfis]

Are you okay with disabling MultiClassesSerializationTest.GpuHist for now?

Go ahead. I will try to reproduce it later.