Fix dynamic inference batching for variable-length tokens

[zjn20030811]

Summary

• right-pad variable-length input_ids, labels, and attention_mask before default collation
• preserve tokenizer, loss-mask, and attention padding values used by the model input preprocessor
• add regression coverage for dynamic token batches and the documented batched inference path

Fixes #67.

Validation

• pytest -q tests/test_serve_policy_dynamic_batching.py (2 passed)
• python -m py_compile src/g05/utils/data/data_utils.py tests/test_serve_policy_dynamic_batching.py
• git diff --check

[zjn20030811]

Implemented and validated the fix for #67.

collate_fn_pad_sequences now right-pads each 1-D token field independently before default_collate: input_ids uses the supplied tokenizer pad id, labels uses -100, and attention_mask uses 0. Equal-length fields and non-token fields remain unchanged. The regression test checks all three pad values, and the existing variable-length batched-vs-sequential inference test passes.

Validation: pytest -q tests/test_serve_policy_dynamic_batching.py (2 passed), py_compile, and git diff --check.