fix: fix DDP race during subtomo update (rank-0-only subtomo rewrite + barrier)

[uermel]

Under multi-GPU DDP, the periodic missing-wedge update had every rank write the
same subtomo .pt files concurrently, truncating them and later crashing the run
with:

PytorchStreamReader failed reading zip archive: failed finding central directory

This PR makes the rewrite rank-0-only and adds a barrier so the other ranks
wait for rank 0 to finish before reading the files again. It also bumps the
DDPStrategy process-group timeout to give the now-serial rewrite headroom.

Cause:

on_train_epoch_end runs on all ranks and calls
update_subtomo_missing_wedges(), which builds a non-distributed DataLoader
over the full dataset (the training DistributedSampler is stripped via .dataset)
and torch.saves every result:

train_set = train_loader.dataset                       # full dataset
dataset   = torch.utils.data.ConcatDataset(datasets)   # full set
loader    = torch.utils.data.DataLoader(dataset, ...)  # no sampler
...
for subtomo, file in zip(subtomo_batch, batch["subtomo0_file"]):
    torch.save(subtomo.cpu().clone(), file)            # N ranks -> same paths

torch.save truncates in place, so N concurrent writers (or a concurrent reader)
corrupt the zip. Training reads are sharded by PL's injected
DistributedSampler; this write path is not — that asymmetry is the bug.