BUG: fix reuse of matfree adjoint Interpolator

[pbrubeck]

Description

Fixes a parallel bug when reusing a matfree adjoint Interpolator. Summation with access=INC skipped the halo exchange after the first call to Interpolator.assemble(). On subsequent calls the summation only happened locally within each process, but not across processes.

This bug did not affect assemble(interpolate(...)) because the Interpolator there is never reused.

[connorjward]

Is there any chance this could be two PRs? It's hard to understand the parallel bits when mixed into the other changes.

Also note that the halo freezing stuff is done automatically in pyop3, so maybe you don't have to worry about optimising that here.

[pbrubeck]

Is there any chance this could be two PRs? It's hard to understand the parallel bits when mixed into the other changes.

I'm having a hard time with git. The parloop removal is the only other thing that's not adressing the bug, should be easy to ignore from the rest of the PR.

[pbrubeck]

@connorjward it turns out that this pointwiseMult is the source of the evil. I fixed it by splitting it in two separate calls, so that multiplication is done in-place.

x.copy(y)
y.pointwiseMult(y, w)

I have no explanation why this approach works and the original doesn't.