Fix parser pad token handling and add loss mask tests

[guojasonliu]

Summary

• Preserve valid tokenizer pad token ids by only falling back to unk_token_id when pad_token_id is None.
• Add parser tests for pad-token handling, assistant loss-mask spans, last_turn_only, and truncation.

Motivation

Aurora builds draft-model training targets from parsed conversation traces. The assistant loss_mask determines which tokens contribute to the training signal. Treating token id 0 as missing can change behavior for tokenizers that validly use 0 as padding.

Testing

• python3 -m pytest tests/test_parse.py -q
• python3 -m ruff check aurora/data/parse.py tests/test_parse.py
• python3 -m ruff format --check aurora/data/parse.py tests/test_parse.py