Move ERI computation to the IPU, keeping sparse symmetric einsum functionality

[mihaipgc]

This PR aims to move ERI computation to the IPU. The code started from nanoDFT/sparse_symmetric_ERI.py and aims to reach a similar form once it is complete.

Overview: 1) sparse_symmetric_intor_ERI.py contains compute_eri which is a reduced version of the relevant integral computing code in electron_repulsion/direct.py

2) compute_eri_utils.py contains minimal modifications to code from electron_repulsion/direct.py to add prescreening

Related issues:

• 55

• 63

Summary of changes:

• [x] correctly computes integral values
• [x] skips all-zero integral subtensors (experimental prescreening)
• [x] supports batching
• [x] maps integral computation to tiles
• [ ] supports multi-ipu integral computation / pmap
• [x] does not store indices (generates them)