Intermittent failure of Windows cmake tests on Internal CI

[gilles-peskine-arm]

On the internal CI only, there are frequent failures in the Windows-20YY-CCC-AAA-cmake* jobs (e.g. Windows-2017-Debug-Win32-cmake, Windows-2013-Debug-x64-cmake-retarget, …). This happens for all branches, both for pull requests and nightly jobs.

Sample log from Windows-2017-Debug-Win32-cmake:
download.txt
It looks like everything is fine until:

   3904:[2024-03-02T02:53:24.520Z]    332>CUSTOMBUILD : CMake error : Cannot restore timestamp C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/tests/CMakeFiles/generate.stamp [C:\builds\workspace\mbed-tls-nightly-tests\worktrees\tmpor0y4a1o\cmake_solution\tests\test_suite_aes.cfb.vcxproj]
   3906:[2024-03-02T02:53:24.520Z]    330>CUSTOMBUILD : CMake error : Cannot restore timestamp C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/tests/CMakeFiles/generate.stamp [C:\builds\workspace\mbed-tls-nightly-tests\worktrees\tmpor0y4a1o\cmake_solution\tests\test_suite_aes.xts.vcxproj]
…
[2024-03-02T02:53:24.521Z]    332>CustomBuild:
[2024-03-02T02:53:24.521Z]          -- Configuring incomplete, errors occurred!
[2024-03-02T02:53:24.521Z]          See also "C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/CMakeFiles/CMakeOutput.log".
[2024-03-02T02:53:24.521Z]          See also "C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/CMakeFiles/CMakeError.log".
[2024-03-02T02:53:24.521Z]    330>CustomBuild:
[2024-03-02T02:53:24.521Z]          -- Configuring incomplete, errors occurred!
[2024-03-02T02:53:24.521Z]          See also "C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/CMakeFiles/CMakeOutput.log".
[2024-03-02T02:53:24.521Z]          See also "C:/builds/workspace/mbed-tls-nightly-tests/worktrees/tmpor0y4a1o/cmake_solution/CMakeFiles/CMakeError.log".

The earliest such failure in nightlies is from 2024-03-01 on 2.28, just after we split the Windows-20NN jobs to parallelize them. But it's plausible that this is a preexisting issue rather than a consequence of this split, because the previous Windows jobs were also failing intermittently. We suspected that these failures were due to a timeout because the job was too long, but the timeout may have been a secondary effect of another problem.