Add secondary LLM selection for chunking

The chunking process can take a while if you're using a larger local model, a lot of 32b models should be able to handle it well, maybe even 14b. We should allow users to select a different model for our condensing logic.

Start by tracing the chunking process and its condensing logic to find where the current model is selected. Determine how users select the primary model, then define the corresponding secondary-model selection and verify that condensing uses it without changing the main model choice.

Written by the indexing model from the issue text.