keep DropOnSync alive until after dispatching for cuda slices

[onursatici]

Summary

cuda device pointers come with SyncOnDrop that is used to synchronise reads and writes to the underlying cuda buffer. We should keep those alive until after dispatching the read or write work on them. For our single stream case this is fine, but if we were to have multiple streams accessing the same buffer these would be a problem

[a10y]

I thought the ones returned from device_ptr don't sync and just record events?

[onursatici]

yes they only record events, but those events are internal to the cuda view and they are used for synchronising other read or write operations on the same slice

[robert3005]

So is there a practical thing to change here or is this something that is useful when combined with another change?

[0ax1]

@onursatici I think this lgtm, but would love to review this tomorrow morning with a fresh mind.

[onursatici]

I believe we don't use the same cuda slice from multiple cuda streams yet so this won't fix anything yet, but it does future proof us for that world