Improved tooltip setup performance

Improved tooltip setup performance

The live collection of `getElementsByClass()` shows a bad performance
when a large list of elements gets remove one-by-one.

It took around 1 second to process 4k elements and stripping the class
from each one. Using `querySelectorAll` instead does the same job in
less than 40ms.

The check against `_elements` is still in place, because it performs
pretty well with large sets of DOM elements (tested with ~16k elements)
and allows us to return early without invoking the costly
`querySelectorAll`, which even without results hit with 0,2 seconds
during tests.