CI: Move msvc to matrix

[dg0yt]

Enable static builds and testing.

... except that the test results must be ignored for Windows: test-timestep_embedding is flaky. The actual result doesn't always match the expected result. Example:

 6/19 Test  #6: test-arange ......................   Passed    0.01 sec
test 7
      Start  7: test-timestep_embedding

7: Test command: D:\a\ggml\ggml\build\bin\Release\test-timestep_embedding.exe
7: Working Directory: D:/a/ggml/ggml/build/tests
7: Environment variables: 
7:  LLVM_PROFILE_FILE=test-timestep_embedding.profraw
7: Test timeout computed to be: 900
7: 0.8439 -0.9970 0.6497 0.9733 0.9981 0.9999 1.0000 -0.5366 -0.0776 0.7602 0.2296 0.0621 0.0167 0.0045 0.0000 0.0000 0.4242 0.9880 -0.1558 0.8946 0.9923 0.9994 1.0000 -0.9056 0.1547 0.9878 0.4470 0.1240 0.0333 0.0089 0.0000 0.0000 
7: -----------------------------------
7: 0.8439 -0.9970 0.6497 0.9733 0.9981 0.9999 1.0000 -0.5366 -0.0776 0.7602 0.2296 0.0621 0.0167 0.0045 242.7236 D:\a\ggml\ggml\tests\test-timestep_embedding.cpp:171: GGML_ASSERT(equalsf(output[i], expected_result[i])) failed
 7/19 Test  #7: test-timestep_embedding ..........Exit code 0xc0000409
***Exception:   0.29 sec

Cf. "Msvc" runs in https://github.com/dg0yt/ggml/actions/workflows/ci.yml.
All errors occured for expected value 0.
Tests didn't run for MSVC before this PR, so the test failure isn't a regression.

I didn't check if there is another way to exclude the flaky test.

[ggerganov]

You can re-enable the failing windows tests so that we have a motivation to fix them. Otherwise we will definitely forget about this.

[dg0yt]

This CI run passed, but the parallel test in dg0yt#3 failed again. Flaky. This will affect other people's PRs here.