Unload model after inactivity

**Is your feature request related to a problem? Please describe.**
Local models use a fair amount of RAM. That’s fine during live transcription, but when the app is idle, it feels wasteful to keep the model loaded.

**Describe the solution you'd like**
Unload the model after some period of inactivity — for example, after 5 minutes with no input.

**Additional context**
In the screenshot, you can see memory usage in Activity Monitor: yellow before closing the app, green after.
Also, when the app first starts, the model isn’t loaded until the dictation key is pressed. That behavior makes sense and ideally should apply after the first dictation too.


Start by tracing the existing lazy model loading triggered by the dictation key in the Swift macOS app. Review how inactivity and model lifetime are currently handled, then verify that the model is released after the requested idle period and loaded again on the next dictation.

Written by the indexing model from the issue text.