Fix SIGBUS in the random test on ARM.

[yuyichao]

We should be robust against non-16bytes aligned addresses but it still has to be naturally aligned.

Fix #12359

[yuyichao]

OSX travis failed with a network timeout during pkg test ..... Restarted.

[Keno]

I don't understand this issue. Why would overalignment be a problem?

[yuyichao]

Maybe I didn't explain this well. Overalignment isn't the problem and the address shift is just for keeping what this test is supposed to be testing.

The test was added in 59e752e to make sure that (AFAICT) certain random functions do not assume all the arrays to be 16bytes aligned. However, the test added create an array that is not aligned to 8bytes for Float64 which causes the bus error on ARM. This PR makes the array aligned to it's natural alignment and still makes sure that we are testing non-16bytes aligned addresses.

[yuyichao]

... sorry I realized that I forgot one line of change when copying the patch over from my test board to my laptop..... Hopefully it makes more sense now (and don't have out of bound array access.....)

[Keno]

Yep, carry on

[yuyichao]

And I include that in the comment now.