Try using multiple workers to improve CI times

Also see if this helps with the OOM errors we've been getting, as the
code generated should be split across multiple workers.

test/runtests.jl

@@ -42,5 +42,20 @@ end
     end
 end
 
-# Run all the LinearAlgebra stdlib tests
-include(joinpath(Sys.STDLIB, "LinearAlgebra", "test", "runtests.jl"))
+# Run all the LinearAlgebra stdlib tests, but with MKL.  We still
+# use `Base.runtests()` to get multithreaded, distributed execution
+# to cut down on CI times, and also to restart workers that trip over
+# the testing RSS limit.  In order for distributed workers to use MKL,
+# we'll modify the test source code so that it imports MKL:
+mktempdir() do dir
+    cp(joinpath(Sys.BINDIR, Base.DATAROOTDIR, "julia", "test"), dir; force=true)
+
+    # Prepend `using MKL` to `testdefs.jl`, so that all test workers load MKL
+    testdefs_content = String(read(joinpath(dir, "testdefs.jl")))
+    open(joinpath(dir, "testdefs.jl"), write=true) do io
+        println(io, "using MKL")
+        println(io, testdefs_content)
+    end
+
+    run(`$(Base.julia_cmd()) --project=$(Base.active_project()) $(dir)/runtests.jl LinearAlgebra`)
+end