Make TargetPointMultiplier a dataclass

[alexfikl]

It has an additional field, so it needs the decorator.

[alexfikl]

FYI, between this and the __reduce__ changes, the Stokes PR is passing locally 🎉

[inducer]

FYI, between this and the __reduce__ changes, the Stokes PR is passing locally 🎉

Oh cool! How come? It looked like the Stokes PR had numerical issues, but these weren't numerical changes? 🤔

[alexfikl]

Oh cool! How come? It looked like the Stokes PR had numerical issues, but these weren't numerical changes? 🤔

As far as I can tell, because this didn't include the axis, it would hash the same for every axis. That would then confuse the OperatorCompiler that would consider IntGs with different kernels to actually be the same.. and then numerical chaos 😁

[inducer]

Oh dear. Good catch! That would do it 🙂. All the more important to dataclass all the things.

[alexfikl]

Oh dear. Good catch! That would do it 🙂. All the more important to dataclass all the things.

Still not quite sure why this only caused issues for the single-layer Stokes / elasticity, but oh well..

[inducer]

Still not quite sure why this only caused issues for the single-layer Stokes / elasticity, but oh well..

It didn't break other parts of pytential because they aren't using this.

[alexfikl]

It didn't break other parts of pytential because they aren't using this.

Yeah, but the Stokes + Laplace + double-layer (in StressletWrapperTornberg) also uses this and it seemed to be converging just fine. I imagine some things were cancelling just right for it..