fix sequence-parallel split for uneven lengths

[zjn20030811]

Summary

• Allow sequence-parallel split-forward communication when the split dimension is not divisible by the process-group size by padding the collective buffer with zeros.
• Keep the Llama hidden states, attention mask, position IDs, and cache positions on the same padded extent, then trim gathered model outputs back to their logical sequence length.
• Align distributed causal-language-model labels with padded local logits and ignore synthetic tail positions in both loss and gradient computation.
• Preserve original extents in the split/gather autograd operations so padded slots cannot leak gradients into valid tokens.
• Add a two-rank CPU gloo regression covering collective values, forward and backward shapes, distributed loss equivalence, and a tiny Llama all-to-all forward/backward pass.

Fixes #6381

Validation

• python -m pytest tests/test_shardformer/test_layer/test_uneven_sequence_split.py -q -s (1 passed)
• python -m py_compile colossalai/shardformer/layer/_operation.py colossalai/shardformer/layer/loss.py colossalai/shardformer/modeling/llama.py tests/test_shardformer/test_layer/test_uneven_sequence_split.py
• ruff check tests/test_shardformer/test_layer/test_uneven_sequence_split.py