Add generate_kvectors cache

[PicoCentauri]

The cache is used within the compute method of FourierSpaceConvolution and is applied when the cell and the number of mesh points has not change compared to the previous call.

However, I did some little tests and the speedup was very minor (~10%). Are you sure @kvhuguenin that generate_kvectors is really responsible for 30% of the total computation time in MeshPotential?

I also renamed smearing to atomic_smearing to consistent within the library and added an example for the usage of FourierSpaceConvolution.

---

📚 Documentation preview 📚: https://meshlode--7.org.readthedocs.build/en/7/

[kvhuguenin]

I have done more systematic tests now. TLDR: everything seems to behave the way it should.

The results are that the savings in computational cost are on the order of 30% without gradients and 15% with gradients for structures with system sizes on the order of N=1e2 to N=1e4. Some comments:

• Firstly, the fact that we save 30% without gradients matches the results from the profiling I had done two weeks ago.
• Furthermore, the fact that we only save half as much with gradients is consistent with the fact that for the backpropagation, all the FFT machinery (the dominant part of the cost) is now run twice, so the relative fraction of the k-vector generation is smaller.
• Also, for structures with N<10, the savings in computational cost are much smaller, most likely because other "constant" contributions in the cost become much more significant.

In summary, the behavior of the caching seems to be exactly what we expect. I just tried to improve the documentation a little, but otherwise, this seems a good addition to the main branch.