Graph breaks in `.forward()` for PMECalculator

[hejamu]

With #219 one can now torch.compile() the whole PME calculations. This already yields a nice speedup for large systems:

PMECalculator, RTX 5080, fp32, bulk-water density (~100 atoms/nm^3)

N	eager	jit.script	torch.compile	torch.compile with reduce-overhead
1000	2.08 ms	1.13 ms	1.33 ms	0.89 ms
4000	2.22 ms	1.27 ms	1.52 ms	1.08 ms
16000	2.52 ms	1.65 ms	1.51 ms	1.07 ms

There is however a large constant caused by graph breaks that reduce-overhead cannot fuse into one CUDA graph because CPU syncs are required.

This is caused by the mesh_interpolator mutating the state in mesh_interpolator.update()

It can be fixed relatively easy by computing the required mesh parameters on the fly and not writing them to self first. On the other hand this is a slight refactor with possible API changes.

I tested this quick and dirty and the numbers look worth the effort:

(Same config as table above)

Code	N	eager	compile(default)	reduce-overhead (CUDA Graph)
PR #219	1000	2.06 ms	1.33 ms	0.89 ms
	4000	2.20 ms	1.51 ms	1.06 ms
	16000	2.54 ms	1.52 ms	1.06 ms
no state mutation	1000	1.82 ms	0.56 ms	0.16 ms
	4000	1.84 ms	0.58 ms	0.23 ms
	16000	1.84 ms	0.57 ms	0.19 ms

I can clean up my changes and open a PR, if this is interesting for others.

[PicoCentauri]

I think it is a good idea. Breaking the API is what we love to do. 😅