Add an optional callback to Lanczos

[RaulPPelaez]

This PR adds a new optional parameter to initialize, a callable that takes an integer and a float.
If passed and lanczos is used to compute the noise the callback will be invoked with the current iteration number and the current error after each iteration.

[github-actions]

✅ Linter reported no issues

All Python files are correctly formatted with Black.

[github-actions]

✅ Linter reported no issues

All C/C++ files are correctly formatted with clang-format.

[rykerfish]

Seems like that did it- for some reason, my thrust libraries are worse than yours so the implicit cpu->gpu copy failed when the CUDA execution policy is specified. Do you think we should add an issue to convert the rest of lanczos to be purely GPU, if that's a feasible task? If so, I'd give it a go sometime in the future.

[rykerfish]

Looks good! My small change fixed the crash I was having so it seems good to go.

[RaulPPelaez]

AFAICT, all the operations I left in the CPU have to be in the CPU. Some decisions must be performed CPU-side (like convergence checking).
I am sure that more performance can be squeezed out of it with some careful profiling, though. If you want to take a go at it, that would be great, for sure.