Automatic classification: more aggressive debounce and perf considerations

Testing #129436 On my machine it takes around `300ms` for the language classification to compute. However some of our users have a slower machine, so here are some ideas on how to make sure to not slow down when users that are just scribbling in an untitled editor. * [x] Do not classify when less than 20 characters present (unless to reset mode if file has auto detected already) - Done in https://github.com/microsoft/vscode-languagedetection/commit/b024b29a080c20b6875141797380dabb1bf3f672 * [x] Increase debounce from 300ms to 600ms (this will still not be a noticeable lag when the user stops typing, pastes) - Done in https://github.com/microsoft/vscode/pull/130006 * [ ] If you have already run the classification `n` times and every time you are unsuccessful maybe stop trying and only on larger changes run * [ ] If you already auto detected a language - great! Do not be so aggressive then, only if a big change to the file content happened Both 3 and 4 require some knowledge about what is a bigger change to a file. We can figure something out. Maybe @bpasero has ideas