fix: keep pretrain validate lockstep on empty ranks (#2319)

[ShamikOfficial]

Summary

• Fixes multi-GPU FSDP deadlock / crash in pretrain.validate when some ranks get zero validation batches (sharded streaming), as reported in litgpt pretrain crashes or deadlocks when a rank's val_dataloader yields zero batches (multi-GPU FSDP) #2319.
• Before each validation step, ranks MIN-reduce a has_batch flag so they stop together; if no rank had data, return NaN instead of torch.stack([]).
• Applies the same lockstep pattern in extensions/thunder/pretrain.py and adds regression tests (empty loader, lockstep protocol, threaded uneven-shard deadlock coverage).

Test plan

• pytest tests/test_pretrain_validate_empty_rank.py -v — 6 passed locally (CPU)
• Empty dataloader returns NaN (no stack([]) crash)
• Uneven shards (rank0 has data, rank1 empty) finish without hanging under a collective-in-forward mock
• Partial overlap (e.g. 3 vs 1 batches) stops after the shared steps only
• Control test: naive empty-skip without MIN-reduce deadlocks as expected
• CI on this PR (GPU / multi-process jobs if available)