Language detection server retains 5Mb if not cleaned up

Had a chat with @TylerLeonhardt on ML code in VS Code and hypothesized that the model might hold on to memory if not cleaned up. Tested on vscode.dev, there is some evidence. In my simple testing, detection did cause a spike to 20Mb but it usually recovered back to 4-7mb, mostly retained in ArrayBuffer/UInt8Array: <img src=https://user-images.githubusercontent.com/8599/135696634-963f731d-23d5-4c39-8710-c86aca197c63.png width=600> Given that it is only needed in special cases and that the memory is nicely contained in a worker that can be quickly stopped and started back up, it would be worth doing some lifecycle management. In Firefox, loading the worker takes about 80ms, while running the model takes 400ms.