Initialize arrays with NaN for easier error detection

[andreasnoack]

The NaNs will make it immediately obvious when some memory hasn't been properly set. I had a case where zero was the correct value so most often the random memory happened to have the correct value. Except on CI where it received memory that made the tests fail.

[tpapp]

Hmm, I was hoping we would catch errors with random tests, but I do see the rationale behind doing something like this.

I am wondering if this would work for float types that have no equivalent of NaN though. This is technically possibly in Julia for a custom <: AbstractFloat type.