feat(models): allow configurable context targets for large-context models

### Problem

Some models offer a large context window, such as 1M tokens, but I do not always need that much conversation history for a task. I would like to use a smaller working context, such as 256K or 512K, to reduce the input tokens carried into later requests.

There is no convenient way to choose this from the conversation controls for a model whose maximum context is already known. Changing the thinking level does not control how much history is retained.

### Desired outcome

- Choose a smaller context target without changing the model's advertised maximum.
- Have Maka compact history earlier when the target is reached.
- Save the setting per connection and model, accessible from both Desktop and CLI.
- Restore automatic behavior when a custom target is no longer needed.
- Keep the conversation controls compact.

The goal is to reduce unnecessary token usage, not to guarantee a fixed cost reduction or impose a hard per-request token limit.

### Alternatives or workarounds

Manual compaction or starting a new task can keep history shorter, but both require intervention and may interrupt the workflow. Editing internal connection configuration is less convenient than a visible control.

AI assistance: OpenAI Codex helped prepare this issue.