Drop the unused NCCL process group from the distributed workers

[clemsgrs]

Summary

The distributed workers build an NCCL process group that has zero live collectives. It can be
removed: read rank/world-size from the environment torchrun already exports, and drop
init_process_group entirely.

Surfaced while scoping distributed dense extraction (the new dense worker deliberately won't use
NCCL); the existing pooled/direct workers don't need it either.

Evidence

• slide2vec/distributed/__init__.py:197 enable() calls
  dist.init_process_group(backend="nccl", timeout=14 days, device_id=...), invoked from
  pipeline_worker.py:32 and direct_embed_worker.py:37.
• Grep across live code (excluding .ipynb_checkpoints): no all_gather, all_reduce,
  broadcast, or reduce. All cross-rank coordination is via the filesystem (shard files / final
  artifacts + the JSONL progress bridge).
• The only dist.barrier() is in hf_login() (slide2vec/utils/config.py:176), reachable only
  from cli.py:27 — the parent process, where is_enabled_and_multiple_gpus() is False. It
  never executes. Workers authenticate via a different path (inference.py:105, plain
  huggingface_hub.login, no barrier).

So the group exists purely so get_global_rank() / get_world_size() can call dist.get_rank() /
dist.get_world_size() — values torchrun already exports as RANK / WORLD_SIZE. Cost: NCCL init,
a comms buffer per rank, and a 14-day-timeout hang surface. Benefit: none.

Latent landmine (fix regardless)

If that barrier ever did run multi-rank with HF_TOKEN unset: non-main ranks hit
if token is None: return and leave, while rank 0 blocks on dist.barrier() alone → a 14-day hang.
Dead today (barrier only runs in the single-process parent); armed the moment anyone calls
hf_login() from a worker. Either remove the barrier or make it collective-safe.

Scope

• Replace dist.get_rank() / dist.get_world_size() in slide2vec/distributed/__init__.py with
  reads of the torchrun-exported env vars; drop init_process_group. Keep torch.cuda.set_device().
• Remove or guard the dead dist.barrier() in hf_login().
• Verify the existing multi-GPU pooled paths (pipeline_worker, direct_embed_worker) still behave
  identically — they never used a collective, so this should be transparent.

References

• slide2vec/distributed/__init__.py (enable, get_global_rank, get_world_size)
• slide2vec/distributed/{pipeline_worker,direct_embed_worker}.py
• slide2vec/utils/config.py:160 (hf_login, the dead barrier)