Investigate numerical errors in `nanoDFT.py`

[AlexanderMath]

The gif below visualizes the numerical distribution of all tensors during our DFT computation.

TODO:

1. Refactor code for plot into utils.py
2. Try to loss scaling, operation pre-scaling/pre-shifting, ... (add HPO and do HPO, i.e., compile one graph, run it 1000s of times exploring options)
3. Add double precision for certain integrals
4. Add decorator @compare which runs graph twice, once in f32 and f64 then compares.
5. ...

[akrzgc]

First point covered here: https://github.com/graphcore-research/pyscf-ipu/pull/118