Refactor the SGD method

[willGraham01]

Related to #75 |

Refactors the stochastic gradient descent method into a standalone function (inside a soon-to-be populated solvers module) that can be called across the code-base.

This has the bonus of making the two_normal_example a bit nicer looking (since we don't have to manually setup the optimisation loop itself).

Other additions include:

• Some basic testing of the new method.
• An implementation of the $l^2$-norm on PyTree objects. This is so that we can take the norms of gradient vectors without worrying about the container that their arguments are passed in as (which in turn dictates the format of the returned gradients).

Main motivation behind doing this is so that we can then apply SGD within the Quadratic Penalty Method, which is one of the focuses of #75.