Fix DDP logging

[edenlightning]

Add a global_zero_only=true flag, if false- create individual files, prefixed with machine nun
Write a logging callback that will do map reduce
Can we do this in the metrics?
Aggregate all tensors on global zero first
(might run into memory issues)
Gather each output individually in CPU memory
We want to preserve the fact that logging is at 0

[edenlightning]

@justusschock

[justusschock]

@edenlightning reducing is something we're currently working on at metrics.

But the logging stuff should be separate I guess. Currently we also gather all results to all devices