Batch cluster provisioning

[peterschmidt85]

Currently, dstack provisions cluster instances one by one. If there is no capacity, it leaves only the instances it was able to provision, which can be fewer than the user requested.

I wonder if there is a better way, such as leveraging the batch provisioning API or using our own algorithm. This requires research because it is unclear what behavior would be most convenient and whether the batch API can solve the problem at all.