Implement logical workers

[voegtlel]

• Allow setting logical_workers in the WorkerConfig. This number may be less than the number of physical workers (=world_size * num_workers), but physical workers must be divisible by logical workers.
• Default: logical_workers = physical_workers (old behavior)
• Example: Setting logical_workers = physical_workers / 4 would result in 4 physical workers iterating the same data, but each with a stride of 4
  • For optimization, one may set @skip_safe or @stateless(skip_safe=True) on functions in the TaskEncoder, resulting in not calling these functions when skipping samples (if the chain supports it).
  • Only functions without fanout and that do not raise errors (to be caught) may be marked as skip_safe.
  • Cannot go "through" packing or filtering (there are non-precomputable (i.e. without reading the sample) fanouts in these).