Chrome's screencast / remote view (in the remote devtools) is apparently *designed* for touch emulation / simulation (https://crbug.com/1410433), which is not convenient when trying to debug mouse-bound issues in a tour (e.g. drag&drop problems).

This can be solved by just running the browser normally if the user wants to watch, in which case the user can also use their normal devtools.

However this doesn't work with the current architecture, where the browser is lazy-started one per test class and reused between `browse_js` calls.

Scrap that, run and shut down the browser entirely for every `browse_js` calls that way it's possible to use entirely different browser. This also simplifies the cleanup phase somewhat as we don't have a persistent profile directory to try and reset, instead every `browse_js` has its own.

Notes:

- `--no-first-run` causes tours to not run at all on my machine, in a non-headless browser
- also moved a bunch of other flags which seem to be mostly for automated annoyances to headless only
- left extensions for now, not entirely sure which is the right one, since we're running with fresh profiles I would assume there's no extensions anyway