Parallelize LinearAlgebra/addmul test

[Keno]

This test does a big loop over various combinations of matrix types: https://github.com/JuliaLang/julia/blob/master/stdlib/LinearAlgebra/test/addmul.jl#L134 and tries to add them together. This test takes the longest of any of the tests in the system. That loop should be fairly easy to distribute across multiple workers. Since each test runs with a different argument combination, I don't think we'd even lose anything by not caching common specializations between workers.

[tkf]

I think we can get even more speedup by launching the workers with --compile=min flag.

[Keno]

Ah, neat idea. We can't do that in general, but it would make sense if we're launching the workers specifically for this test.

[andreasnoack]

I think we can get even more speedup by launching the workers with --compile=min flag.

If this become possible, we could probably reduce CI time for LinearAlgebra by a lot. E.g. almost all of the time in triangular.jl is spent on compilation.

[ViralBShah]

Do we still need to do something here?