Add synchronizes before/after managed_ptr kernels

[adayton1]

managed_ptr kernels currently run in stream 0, which is different than the default stream used by RAJA. To prevent any cross stream race conditions, we synchronize before and after those kernels. In the future, we will likely use the same default stream used by RAJA and/or modify the API such that there is an overload that takes a camp resource or stream as an argument.

[robinson96]

LGTM

[MrBurmark]

I thought the streams created by camp should be synchronous with stream 0.

[adayton1]

I thought the streams created by camp should be synchronous with stream 0.

Is that true for both cuda and hip?

[MrBurmark]

I know its true for cuda, and I assume its true for hip, but I'll check.

[adayton1]

I know its true for cuda, and I assume its true for hip, but I'll check.

I can't find anything concrete, but the hip documentation seems to hint at it.

EDIT: I found slides that seem to indicate it has the same behavior as CUDA. "No task on the NULL stream will begin until all previously enqueued tasks in all other streams have completed."

But this leads me to the question of whether tasks on other streams can start while tasks are running in the null stream.