fix: explicitly stop models to prevent concurrent loading errors

[marcelrv]

Explicitly calls ollama stop after each run to ensure memory is freed. This prevents the benchmark from failing on systems where Ollama attempts to load multiple models simultaneously, which otherwise leads to resource exhaustion and prevents completion.

[chuangtc]

The fix is safe to merge if your ollama is reasonably recent (≥0.1.24). But it would be more robust with check=False or a try/except.