Thinking Machines Launches Tinker: A Low-Level Training API that Abstracts Distributed LLM Fine-Tuning without Hiding the Knobs