fix: add ollama_num_ctx for longer context length

[anxkhn]

Currently, when using Ollama inference, it defaults to num_ctx=2048, which is too small for longer prompts and outputs. This results in truncated responses, as confirmed by the Ollama logs.

This PR updates our configuration and utility code to set num_ctx explicitly, preventing prompt and response cutoffs for detailed posts. Following Ollama’s guidance, we set num_ctx to a higher value (default 8192 - seems like a sweet spot ensuring compatibility with smaller machines), and Ollama will automatically cap it at the model’s maximum supported context size.

Changes:

• Added ollama_num_ctx setting to config.toml (default: 8192).
• Updated ollama_predict in leetcomp/utils.py to pass num_ctx from config.

Result:
Ensures longer prompts and responses are not truncated, without requiring per-model manual tuning.