Optionally allow allocations to be zeroed

[llnl]

Optionally allow allocations to be zeroed

[brunner6]

Here's a comment from @rchen20 in our internal issue:

I don't think there are any plans to add zero-ing memory to RCU, other than what is being used in this MR (Umpire memset). Looking at the functor implementation of kConst, I would guess it would be faster than umpire::memset. I can bring this up in the next RAJA or Umpire meeting; maybe we can propose adding the Armus::kConst implementation to Umpire?

And here's a follow-up from @ajkunen :

I think this functionality should go in CHAI:

• RAJA doesn't do any memory management
• One could make an Umpire allocator that zeros new allocations... but you really just want the first allocation to be zeroed, not both the host and device allocations.
• CHAI is the only thing that knows that host and device allocations are associated. So, CHAI would be the right thing to trigger the zeroing... but I think Umpire should do the zeroing (like @rchen20 suggests)... so CHAI should tell Umpire to zero the first allocation for a ManagedArray, on whatever device that happens to be.

We could get this functionality now, by wiring this logic into our existing CHAI callback function... but it might be cleaner to just let the CHAI team implement it.