fix: accept token lists in parse results

[suguanYang]

Summary

• fix TextChunk.tokens deserialization so the SDK matches the current knowhere-api list-based token payload
• remove the old integer/string compatibility path and fail fast if chunks.json contains an outdated token shape
• update parser/model/docs/tests and refresh the regression fixture to the current response contract

Context

Issue #7 reports that client.parse() crashes after a successful parse job when a text chunk returns tokens as a list instead of an integer. Local inspection of recent knowhere-api changes shows list-based tokens are now the intended backend contract.

Closes #7.

Testing

• uv run pytest
• uv run pytest tests/test_result_parser.py tests/test_models.py