Consider thawing loop in same distributed map as copy

[andrewpatto]

The thawing steps is currently in its own distributed map - which waits for all thawing to finish (across all files). It then proceeds to the copy phase which is its own distributed map. This was done because the scale-out factor of the two are different (one is scaling out thousands of tiny lambdas to do thawing, the other is scaling out a limited set of fargate jobs).

We could consider refactoring this

• so that objects that are already thawed can copy straight away
• copies could immediately proceed after thawing without waiting for the slowest