docs: Fit models with standard for loop + notebook on training loops.

[daniel-dodd]

We have a convenient fit function to train GPs against objectives. It would be good though e.g., in the regression notebook to show a simple python for loop, and a simple lax.scan training loop, to demonstrate that users can write their own training loops. Give insight to, eg., the ConjugateMLL is something you can take jax.grad against and just do gradient descent on.

Then it would be good to link this to a more extensive notebook exposing users to stoping gradients, bijectors transformations etc, and show how to add a training bar to the loop.