Modify docs to emphasize faster pattern for Python thread pools

[itamarst]

Having determined that the time spent loading libraries can add up in real-world, it's worth only documenting the faster pattern for Python thread pools:

1. Create one ThreadpoolController
2. Re-use it in the threads

It may also be worth documenting while I'm at it the limits of having a single ThreadpoolController: not noticing the existence of new libraries, usually due to later imports.

[itamarst]

(I will also update the relevant joblib branch and future PR to use this pattern.)